Artificial intelligence writing 90% of a company’s software sounds like the beginning of a mass layoff announcement. Anthropic CEO Dario Amodei sees it differently—at least initially. In his view, automating most of a job does not immediately eliminate the worker. It creates a productivity surge in which humans concentrate their time on the small portion the machine still cannot complete.

That distinction sits at the center of Amodei’s increasingly provocative argument about the future of work.

When Claude generates most of the code, engineers do not necessarily disappear. They become reviewers, architects, product designers, security investigators and managers of increasingly capable digital workers. The human contribution shrinks as a percentage of the production process, but the output of each person can rise dramatically.

The more unsettling question is what happens when AI masters the final 10%.

Amodei’s answer reaches far beyond software development. He imagines artificial intelligence becoming the cognitive core of companies, helping organizations reason, coordinate and execute at a level that makes the modern enterprise resemble a form of collective superintelligence.

It is an ambitious vision combining extraordinary productivity, accelerated medical discovery and potentially severe disruption to white-collar employment.

Writing Code Is Not the Same as Doing the Job

The percentage of code written by AI has become one of the most widely repeated statistics in the technology industry.

Amodei has said that Claude now produces most of the code written by some engineers inside Anthropic. In parts of the company, developers may no longer type significant amounts of code manually. They describe the intended feature, direct the model, inspect its output, test the implementation and intervene when something goes wrong.

This is a fundamental change in the interface between an engineer and a computer.

Traditional software development requires humans to translate ideas into precise instructions written in programming languages. AI coding agents can absorb much of that translation work. A developer can increasingly communicate at the level of goals, constraints and architecture while the model handles implementation.

But lines of code are a poor measurement of complete job automation.

Compilers already generate enormous quantities of machine code, yet their arrival did not make programmers unnecessary. Higher-level programming languages automated much of the work once performed manually, allowing developers to build larger and more complex systems.

Claude writing 90% of a codebase may therefore say less about the disappearance of engineers than it does about the abstraction level at which they work.

The remaining 10% can still contain the most difficult and consequential decisions. Someone must determine what should be built, understand the needs of users, choose between competing technical designs, identify security risks and decide whether the output is safe to deploy.

AI can generate a plausible implementation in minutes. Knowing whether it solves the correct problem remains a different challenge.

The Productivity Hump

Amodei describes a transitional period in which automation produces an enormous increase in productivity before it produces full replacement.

Imagine that AI can reliably perform 90% of the work involved in a software project. The engineer is still necessary because the final 10% requires human judgment, organizational knowledge or technical expertise. Yet the engineer can now spend nearly all available time on those remaining tasks.

In simplified terms, one engineer may become capable of supervising the volume of work previously handled by ten.

Companies could respond by reducing staff, but they could also build far more software. Projects previously rejected as too expensive could become viable. Internal tools that never reached the top of the development queue could be created quickly. Small teams could launch products that once required large engineering departments.

This is the productivity hump: the period in which humans remain essential but become dramatically more leveraged.

The economic consequences will depend on how much additional demand appears. When productivity rises, companies do not always reduce employment proportionally. Lower costs can create new markets, new products and new categories of work.

However, that protection has limits.

If AI advances from writing most of the code to completing nearly the entire software-engineering process, the remaining human bottleneck begins to disappear. The model would not merely implement a feature. It would identify the requirement, inspect the existing system, design the solution, configure the environment, run tests, diagnose failures, document the change and prepare it for deployment.

At that point, engineering becomes less about humans using better tools and more about humans assigning objectives to autonomous systems.

From Roughly 5% to More Than 77%

The speed of improvement in coding benchmarks helps explain Amodei’s confidence.

The original SWE-bench evaluation was designed around genuine software issues collected from public GitHub repositories. Instead of asking a model to write a small function or solve an interview-style coding puzzle, it required the system to understand an existing codebase and generate a patch that resolved a documented problem.

Early results were poor. Claude 2 resolved only a small percentage of the tasks under the initial evaluation setup. The result demonstrated how far language models still had to go before they could perform practical repository-level software engineering.

Later Claude models made rapid gains. Anthropic reported that Claude Sonnet 4.5 achieved 77.2% on SWE-bench Verified, a human-reviewed subset containing 500 software problems.

The figures should not be treated as a perfectly controlled comparison. The benchmark variant, model scaffolding, prompting strategy, tool access and evaluation methodology changed over time. A score on the original benchmark is not directly interchangeable with a score on the Verified subset.

Even with those caveats, the direction of travel is difficult to ignore.

AI coding systems have moved from solving only the simplest isolated issues to handling substantial portions of carefully selected real-world software tasks. They can navigate repositories, edit multiple files, execute commands, run tests and revise their own attempts.

Benchmarks still do not capture the complete reality of production engineering. Real companies have undocumented systems, conflicting stakeholder demands, legacy infrastructure and security requirements that cannot be represented by a clean test suite.

Yet the improvement suggests that the islands of work reserved for humans are becoming smaller.

A Medical Story With Larger Implications

Amodei has also used a personal family experience to illustrate how AI can identify patterns across complicated information.

According to his account, his sister and Anthropic co-founder Daniela Amodei developed an infection while pregnant. Several doctors believed the illness was viral. After her medical information was provided to Claude, the model suggested that the infection could instead be bacterial.

The anecdote is powerful because it captures a potential advantage of medical AI: the ability to process a large volume of records, symptoms and reference material without fatigue.

A doctor may have limited time with each patient and may receive information spread across laboratory reports, previous appointments, medication histories and specialist notes. A model can examine those records together and surface possibilities that deserve another look.

That does not make Claude a replacement for a physician.

A personal account is not a clinical trial, and an AI-generated suggestion should not be treated as a verified diagnosis. Language models can misunderstand records, overlook critical context or produce confident but inaccurate conclusions. Medical decisions also require physical examinations, professional accountability and an understanding of the patient that cannot always be captured in uploaded data.

The more realistic near-term role is that of a second reader.

An AI system can summarize a patient’s history, identify unusual combinations of symptoms, compare test results over time and suggest questions for a clinician. The doctor remains responsible for evaluating those suggestions and deciding whether further tests or treatments are appropriate.

The same productivity dynamic seen in coding could emerge in medicine. AI handles the information-intensive portion of the work, allowing medical professionals to spend more time on difficult judgments, procedures and patient relationships.

The stakes, however, are much higher. A coding error may break an application. A medical error can harm a person.

The Enterprise as a Collective Intelligence

Amodei’s broadest idea concerns the nature of the company itself.

An enterprise already behaves like a distributed intelligence. It collects information from customers and markets, stores institutional knowledge, assigns tasks, makes decisions and coordinates the actions of thousands of people.

Executives act as strategic planners. Managers distribute information and resources. Employees operate as specialized units. Databases and software systems function as organizational memory.

The result is more capable than any individual person.

Placing AI at the center of that structure could make the organization faster, more coordinated and more responsive. Instead of acting as a chatbot used by isolated employees, the model could become a shared reasoning layer connected to the company’s data, applications and operational processes.

An AI-centered enterprise might monitor sales activity, examine customer feedback, analyze product performance and recommend changes continuously. It could draft software updates, prepare financial forecasts, identify supply-chain risks and coordinate specialized agents responsible for different departments.

Human employees would establish objectives, approve sensitive decisions and intervene when judgment or accountability is required.

In this model, AI is not simply another application purchased by the information-technology department. It becomes part of the company’s operating system.

That prospect explains why enterprise AI is strategically important to Anthropic. Consumer chatbots attract public attention, but organizations control enormous collections of proprietary data and repeatable workflows. Connecting models to those systems could generate far greater economic value than answering standalone questions.

The New Bottleneck Is Judgment

As AI takes over execution, the value of human work may shift toward deciding what deserves to be executed.

A model can write a technically correct feature that customers do not need. It can optimize a metric that damages the wider business. It can confidently follow instructions that were badly designed from the beginning.

Greater execution capacity can therefore magnify poor judgment.

When software becomes cheaper to produce, companies may generate more unnecessary complexity. When reports become effortless to create, employees may drown in synthetic analysis. When autonomous agents can perform thousands of actions, a poorly specified objective can produce failures at extraordinary speed.

The most valuable workers may be those who understand systems deeply enough to direct AI effectively and recognize when its output is misleading.

That requires more than clever prompting. It requires domain knowledge, skepticism, taste and accountability.

Junior roles present a particular challenge. Companies traditionally develop senior experts by giving beginners routine tasks and gradually exposing them to harder problems. If AI absorbs the entry-level work, organizations may struggle to train the people eventually expected to supervise advanced systems.

A company cannot indefinitely remove the bottom rung of the career ladder while expecting experienced professionals to appear at the top.

Productivity and Displacement Can Both Be True

The optimistic and pessimistic interpretations of Amodei’s argument are not mutually exclusive.

AI can make engineers ten times more productive and still reduce the total number of engineers companies need. It can create new products while eliminating familiar roles. It can help doctors detect overlooked conditions while introducing new forms of diagnostic risk.

The outcome will not be determined by a single automation percentage.

It will depend on how quickly new demand develops, whether organizations reinvest productivity gains, how governments respond and whether humans can continue moving into new areas of comparative advantage.

The transition may also unfold unevenly. The strongest engineers could become dramatically more valuable because they can manage fleets of coding agents. Less experienced developers may face fewer opportunities. Large companies could become leaner, while small teams gain the power to compete with established organizations.

The result could be both democratizing and concentrating at the same time.

What Happens When AI Learns the Rest?

The most important part of Amodei’s argument is not that Claude writes 90% of the code. It is that the remaining percentage may not remain protected for long.

Today’s models still need supervision. They make mistakes, lose track of objectives and struggle with ambiguous organizational realities. Humans remain necessary because the final portion of the task contains uncertainty, responsibility and context.

But frontier AI companies are specifically working to improve reasoning, memory, tool use and long-horizon autonomy—the capabilities required to attack that final portion.

The productivity hump may therefore be temporary.

For now, AI allows one person to accomplish far more. The engineer becomes an architect. The doctor gains a tireless second reader. The enterprise acquires a new layer of collective intelligence.

Beyond that stage lies a harder question: not how humans work with machines, but what economic role remains when machines can carry an objective from conception to completion.

Amodei’s vision is compelling because it contains both possibilities. AI could become the greatest amplifier of human capability ever created. It could also advance so quickly that the new roles it creates are automated almost as soon as people learn to perform them.

The decisive battle will not be over the first 90%.

It will be over the last 10%.

The most important improvement in GPT-5.6 Sol is not that it can produce a sharper answer to a difficult question. It is that the model is increasingly capable of turning an ambiguous objective into a sequence of actions, carrying those actions across tools, checking the results and returning something that resembles finished professional work. That distinction matters because the artificial intelligence market is moving beyond the chatbot era. The next competitive frontier is not conversation. It is execution.

Released for general availability on July 9, 2026, GPT-5.6 Sol sits at the top of OpenAI’s new three-tier model family. Sol is the flagship, Terra balances capability with cost, and Luna is optimized for speed and affordability. The naming change is more than branding. It reflects an industry-wide shift away from presenting each model as a single, static intelligence and toward selling families of systems that can allocate different amounts of reasoning, computation and agent activity depending on the task.

Sol is therefore best understood as a professional execution engine. It is designed for software engineering, research, cybersecurity, scientific analysis, document creation, computer use and other workflows in which the model must maintain context, operate tools and revise its own work. It also introduces OpenAI’s most ambitious multi-agent mode so far, allowing several coordinated model instances to investigate different parts of the same problem in parallel.

The result is one of OpenAI’s most consequential releases since the company began turning general-purpose language models into operational agents. GPT-5.6 Sol does not win every benchmark, nor does it eliminate the strengths of Claude, Gemini, Grok or DeepSeek. Its more significant achievement is combining frontier-level reasoning with a serious attempt to control the cost, latency and token consumption of autonomous AI work.

What GPT-5.6 Sol Actually Is

GPT-5.6 is a family rather than a single model. Sol occupies the premium capability tier, roughly replacing the role played by the unsuffixed flagship models in previous GPT generations. Terra is positioned as the practical middle option, while Luna targets high-volume applications where response speed and operating cost matter more than extracting the final percentage points of intelligence.

For developers, GPT-5.6 Sol supports text and image input and produces text output. Its API specification provides a context window of approximately 1.05 million tokens and a maximum output length of 128,000 tokens. That gives the model enough theoretical capacity to inspect enormous codebases, extensive legal or financial records, long research collections and complex multi-document projects within a single working context.

A large context window, however, is only useful when the model can identify and preserve the right information. Frontier models have repeatedly demonstrated that accepting a million tokens is not the same as reasoning reliably across a million tokens. Sol shows substantial improvements on several long-context tests, but its results are not uniformly dominant. On some evaluations involving very large contexts, competing Claude models remain highly competitive, and GPT-5.5 occasionally matches or narrowly exceeds Sol.

The more important improvement is therefore not raw context size. It is how Sol combines context with reasoning, tool use and iterative execution. The model can write lightweight programs to process intermediate data, coordinate tools and decide what to do next. Instead of repeatedly sending every tool result back through a conventional conversational loop, it can filter and transform information programmatically, retaining only what is useful for the next step.

This is a major architectural shift at the product level, even though OpenAI has not disclosed every detail of the model’s underlying neural architecture. The system is being optimized around completed workflows rather than isolated responses.

From GPT-5.5 to GPT-5.6: A Change in Operating Philosophy

GPT-5.5, released in April 2026, already represented a substantial move toward agentic work. It was designed to understand messy requests, navigate software, use external tools, research information and continue working without requiring the user to supervise every decision. GPT-5.6 Sol extends that direction but places much greater emphasis on efficiency, parallelism and polished output.

The difference can be seen in how the two generations approach complexity. GPT-5.5 was a stronger autonomous worker than GPT-5.4, particularly in coding, computer use and document-heavy tasks. Sol is designed to make that worker more economical and more adaptable. It can invest additional reasoning only where it is likely to improve the result, while using fewer tokens on routine stages of the workflow.

That distinction becomes significant at enterprise scale. A model that solves a task 5 percent more accurately but consumes twice as many tokens may be unsuitable for production. It may also become slower as the workflow expands, particularly when an agent repeatedly reads large tool outputs, revisits previous reasoning or generates unnecessary explanations. OpenAI’s emphasis on token efficiency suggests that the company increasingly views wasteful inference as a product defect rather than an unavoidable cost of higher intelligence.

The performance differences are especially visible in computer use and cybersecurity. OpenAI’s published evaluations show Sol making large gains over GPT-5.5 on operating-system tasks, browsing, computer-aided design and security research. The improvement in general academic reasoning is more incremental. Sol scores above GPT-5.5 on demanding science and mathematics evaluations, but the gap is smaller than it is on tasks requiring sustained interaction with tools.

This pattern reveals the real purpose of the release. GPT-5.6 is not primarily a better examination candidate. It is a better operator.

Reasoning That Can Scale Up When Necessary

Sol introduces several levels of reasoning effort, allowing users and applications to choose between faster execution and deeper analysis. The new “max” setting gives the model more time to explore alternatives, test assumptions and revise its approach than the previous highest reasoning configurations.

The more dramatic feature is “ultra,” which moves beyond a single reasoning process. In its default configuration, ultra coordinates four agents operating in parallel. Different agents can research separate questions, test competing approaches or perform independent checks before a root agent synthesizes their results.

Multi-agent systems are not automatically superior. Four agents can consume more tokens, duplicate effort or amplify the same incorrect assumption. Coordination itself can become a source of failure if the agents divide the problem poorly or if the final synthesizer cannot distinguish strong evidence from confident noise.

OpenAI’s implementation is therefore important because it treats multi-agent reasoning as an optional escalation mechanism rather than the default response to every prompt. Routine tasks can remain on a lower reasoning setting, while research, engineering and strategic analysis can receive additional computational investment.

This resembles how professional teams allocate human effort. A straightforward memo does not need four analysts. A complex acquisition, software migration or security investigation might. The advantage is not simply having more intelligence. It is being able to match the amount and organization of intelligence to the economic value of the task.

Ultra also changes the relationship between latency and capability. Parallel agents may use more total tokens, but they can complete independent workstreams simultaneously. For time-sensitive projects, the result may arrive faster than a single agent working through every branch sequentially. The trade-off is a higher total inference bill in exchange for greater breadth, stronger cross-checking and a shorter time to completion.

Coding Becomes a Full Engineering Workflow

Coding remains one of Sol’s strongest areas, but describing it as a code-generation model would understate the change. The model is designed to operate across the engineering lifecycle: inspecting repositories, understanding architecture, reproducing failures, editing files, running tests, reviewing results and continuing until the implementation works.

On OpenAI’s reported Terminal-Bench 2.1 evaluation, Sol reaches 88.8 percent, while ultra rises to 91.9 percent. GPT-5.5 records 85.6 percent in the same comparison. Sol also improves on long-horizon engineering tests involving real codebases and command-line environments.

Those gains are meaningful because terminal benchmarks are harder to game with elegant-looking but nonfunctional code. The model must use tools, cope with errors and maintain a plan across multiple actions. This is closer to the way engineering work actually happens.

The results are not a universal victory. On SWE-Bench Pro, Anthropic’s Claude Fable 5 and Mythos 5 configurations score substantially higher than Sol in OpenAI’s own comparison table. That makes Claude a formidable option for resolving difficult repository issues, especially when long autonomous runs and codebase comprehension are central to the task.

Sol’s case rests on the wider workflow. It combines strong coding performance with computer use, artifact creation, programmatic tool coordination and lower list pricing than Anthropic’s top models. A company choosing between Sol and Fable may therefore reach different conclusions depending on whether it needs the highest success rate on a narrow software benchmark or a versatile agent that moves between code, research, files, interfaces and presentation-ready deliverables.

For crypto companies, the potential applications are obvious but should be approached carefully. Sol can assist with smart-contract review, transaction-analysis pipelines, test generation, protocol documentation and incident investigation. It can also accelerate dangerous security work, which explains why access to some cyber capabilities is governed by stricter safeguards. No serious team should treat model-generated security analysis as a substitute for independent audits, deterministic testing and human review.

Knowledge Work Moves From Drafting to Delivery

Earlier generations of generative AI were useful for producing first drafts. They could summarize a report, outline a presentation or suggest spreadsheet formulas, but the user usually had to transform the output into a finished artifact.

GPT-5.6 Sol aims to reduce that final-mile burden. It can take unstructured information from documents, workplace messages, cloud drives and productivity platforms, then turn it into reports, financial models, presentations and other editable outputs. OpenAI places particular emphasis on Sol’s ability to follow existing templates, infer visual systems and preserve recurring design conventions.

This may sound cosmetic, but formatting is part of professional accuracy. A model that produces correct analysis but ignores a company’s slide master, omits required sections or breaks a financial template has not finished the job. It has merely transferred the remaining work to a human.

Sol’s stronger design judgment is therefore strategically relevant. It can inspect rendered output rather than focusing only on the underlying code or text. In practical terms, this means checking whether a page is visually coherent, whether an interface is usable or whether a presentation follows the reference material.

OpenAI’s evaluations show significant gains over GPT-5.5 on browsing, computer use and computer-aided design. Sol reaches 62.6 percent on OSWorld 2.0 compared with 47.5 percent for GPT-5.5. It scores 70.6 percent on BenchCAD compared with 44.4 percent for its predecessor. Sol Ultra reaches 92.2 percent on BrowseComp, while standard Sol records 90.4 percent and GPT-5.5 reaches 84.4 percent.

The broader benefit is not simply higher quality. It is fewer revision cycles. In enterprise deployments, every additional prompt, correction and manual handoff adds cost. A model that understands the expected format and validates its own output can create value even when its raw reasoning score is only modestly higher.

The Economics of Token Efficiency

Sol is priced at $5 per million input tokens and $30 per million output tokens through the OpenAI API. Terra costs $2.50 for input and $15 for output, while Luna costs $1 and $6 respectively. Cached input for Sol receives a substantial discount, although the GPT-5.6 family also introduces a charge for writing new cache entries.

These prices make Sol expensive compared with high-volume models such as Gemini 3.5 Flash, but relatively economical compared with Anthropic’s Claude Fable 5, which is listed at $10 per million input tokens and $50 per million output tokens.

Token pricing alone does not reveal the real cost of a workflow. A cheaper model may produce a longer answer, require more retries or fail often enough that the effective cost per successful task becomes higher. An expensive model can be economical when it completes difficult work on the first attempt.

OpenAI is explicitly positioning Sol around this idea of performance per dollar. The company claims that Sol uses fewer output tokens, less time and lower estimated cost than several competing frontier models on selected agentic evaluations. Even where Sol does not lead the raw intelligence score, it may reach a similar result with less computation.

This is one of the most important changes in the AI market. Model buyers are becoming less interested in price per token and more interested in cost per completed outcome. A legal team does not buy tokens; it buys reviewed contracts. A software company buys resolved issues. A financial institution buys validated analysis. An AI model that generates millions of cheap tokens without completing the workflow can be more expensive than a premium system that finishes accurately.

Sol’s efficiency narrative will need independent validation under real production conditions. Vendor estimates may not account for every tool call, failure mode, latency spike or integration expense. Nevertheless, the focus is correct. The next stage of AI adoption will be determined by unit economics as much as benchmark intelligence.

GPT-5.6 Sol Versus Claude Fable 5 and Mythos 5

Anthropic remains Sol’s most direct competitor for demanding professional and coding tasks. Claude Fable 5 is Anthropic’s most capable generally available model, while Mythos 5 uses the same underlying model with different safeguards and restricted access for selected cybersecurity and scientific users.

Fable 5 is particularly strong on long-running autonomous work, software engineering, vision, finance and scientific research. Anthropic says the model can sustain attention across millions of tokens and use persistent notes to improve performance over extended tasks. Early customers have reported impressive results on codebase migrations, legal review, analytics and research.

OpenAI’s own evaluations present a mixed but revealing comparison. Sol leads Fable on Agents’ Last Exam and on the Artificial Analysis Coding Agent Index. It also achieves stronger results on Terminal-Bench 2.1. Fable, however, substantially outperforms Sol on SWE-Bench Pro and narrowly leads on the broader Artificial Analysis Intelligence Index. Claude configurations also outperform Sol on Toolathlon, an evaluation of complex tool use.

The pricing difference favors OpenAI. Sol’s standard API rates are half of Fable’s input price and 40 percent lower on output. OpenAI also claims major advantages in latency and token usage on selected tasks.

Claude’s appeal is not limited to benchmarks. Many users prefer its writing style, long-form coherence and measured handling of complicated documents. Anthropic has also built a strong reputation among developers through Claude Code and integrations with engineering platforms. Fable may remain the preferred option for teams that prioritize autonomous repository work, nuanced writing or exceptionally long research sessions.

Sol is the stronger choice when the workflow crosses more boundaries. It is designed to move naturally between research, coding, computer interaction, visual design and structured artifact generation. The competition is therefore not a simple question of which model is smarter. Fable resembles a highly capable specialist with exceptional endurance. Sol resembles a versatile operating layer built to coordinate an entire digital project.

GPT-5.6 Sol Versus Google Gemini

Google’s competitive position is different because Gemini is connected to one of the world’s largest software and data ecosystems. Gemini models can be integrated across Search, Workspace, Android, Google Cloud and enterprise agent platforms. That distribution can matter more than a narrow benchmark victory.

As of Sol’s launch, Google’s most widely deployed new model is Gemini 3.5 Flash. Despite the Flash label, it is positioned as a frontier-level agentic and coding model rather than a lightweight assistant. Google reports strong results on Terminal-Bench, multimodal reasoning and agentic workflows, with high output speed and built-in computer-use capabilities.

Gemini 3.5 Flash costs $1.50 per million input tokens and $9 per million output tokens, making it significantly cheaper than Sol. It is therefore attractive for high-volume agents, customer-facing systems, search-based applications and workflows where latency matters more than maximum reasoning depth.

Sol has the advantage on several of OpenAI’s reported professional, coding and scientific evaluations. It also offers max and ultra reasoning for tasks that justify additional computation. Gemini’s strategic advantage lies in multimodality, speed, global distribution and direct access to Google’s product ecosystem.

Gemini 3.1 Pro remains relevant for deeper reasoning comparisons, although Google has been transitioning attention toward the 3.5 generation. In OpenAI’s published tables, Sol substantially outperforms Gemini 3.1 Pro Preview on coding, professional work, browsing and several science evaluations. Gemini remains close on multimodal academic reasoning and benefits from Google’s experience with video, audio, search and large-scale infrastructure.

For enterprise buyers, the decision may be shaped by where their data already lives. An organization centered on Google Cloud and Workspace may prefer Gemini even when Sol has a benchmark advantage. The integration cost, identity system, governance structure and data permissions can outweigh small differences in model quality. Sol’s challenge is to be sufficiently better at completing work that companies accept the cost and complexity of adding another AI platform.

GPT-5.6 Sol Versus Grok 4.5

Grok 4.5, released one day before GPT-5.6’s general launch, is SpaceXAI’s strongest model for coding, knowledge work and agentic tasks. It is designed for fast inference and deep integration with engineering tools, including Cursor and Grok Build.

SpaceXAI reports that Grok 4.5 is served at around 80 tokens per second and uses far fewer output tokens than Claude Opus 4.8 on selected software-engineering tasks. It also performs competitively on Terminal-Bench 2.1, although OpenAI’s newer Sol results exceed the Grok scores published at launch.

Grok’s differentiator is its connection to real-time search and the X platform. The base model does not automatically know current events beyond its training cutoff, but developers can add web and X search tools. This can make Grok attractive for live market monitoring, public-sentiment analysis, news tracking and fast-moving research.

Those capabilities are particularly relevant in crypto, where narratives, token flows, governance disputes and market reactions evolve continuously. A Grok-based system can monitor public conversation and breaking developments, while a Sol-based agent may be better suited to converting that information into a structured investment memo, analytical model, codebase or operational plan.

Grok also competes through speed and a more permissive product identity. Sol competes through broader professional execution, stronger reported computer use, mature artifact generation and a larger enterprise productivity ecosystem through OpenAI and Microsoft.

The contest is still early. Grok 4.5’s launch information does not provide enough standardized data for a definitive head-to-head judgment against Sol. What is clear is that SpaceXAI is no longer competing only on personality or access to X. It is targeting the same valuable engineering and agentic workloads as OpenAI and Anthropic.

GPT-5.6 Sol Versus DeepSeek V4

DeepSeek V4 represents a different kind of pressure. It is not merely another proprietary chatbot. It is an open-weight model family designed to offer strong reasoning and agent capabilities at dramatically lower infrastructure and API costs.

The V4 family includes a Pro model with 1.6 trillion total parameters and 49 billion activated parameters, as well as a smaller Flash version with 284 billion total parameters and 13 billion activated. Both support contexts of approximately one million tokens. Because they use a mixture-of-experts design, only part of the model is activated for each token, improving inference efficiency.

DeepSeek’s strategic advantage is control. Organizations can inspect, modify and self-host open models, subject to licensing and technical constraints. This is valuable for governments, research institutions, crypto protocols and companies that cannot send sensitive data to an external proprietary API.

Sol is likely to be easier to deploy for teams that want a polished managed service, integrated tools, strong multimodal input and enterprise support. DeepSeek is more appealing for organizations willing to invest in infrastructure in exchange for customization, data sovereignty and lower marginal cost.

The current V4 release is also a preview, and open deployment brings its own burdens. Hosting a trillion-parameter mixture-of-experts model is not a casual undertaking. Teams must handle hardware, optimization, monitoring, security, model updates and reliability. “Open” does not mean operationally free.

DeepSeek’s presence nevertheless changes the market. It prevents frontier AI from becoming a competition only among premium American APIs. Even when Sol delivers better overall performance, DeepSeek can force OpenAI to defend its pricing and offer clearer economic value. The more capable open models become, the less customers will tolerate paying a large premium for intelligence that does not produce a correspondingly better business result.

Cybersecurity Is Both a Benefit and a Constraint

GPT-5.6 Sol delivers some of its largest improvements in cybersecurity. On OpenAI’s ExploitBench comparison, Sol scores 73.5 percent against GPT-5.5’s 47.9 percent. On SEC-Bench Pro, it reaches 71.2 percent compared with 45.8 percent for GPT-5.5. Its ExploitGym performance also more than doubles the predecessor’s result under the longest published evaluation period.

These capabilities can help defenders review code, identify vulnerabilities, develop patches, perform threat modeling and analyze malware. They can also lower the skill required to conduct harmful attacks.

OpenAI has responded with a layered safeguard system combining behavior trained into the model, real-time monitoring, account-level signals and access controls. The company says Sol blocks far more potentially dangerous cyber activity than previous models and reserves some advanced defensive capabilities for verified users.

The downside is increased friction. Legitimate security researchers may encounter refusals, additional checks or requests that are redirected to less capable models. OpenAI acknowledges that its initial approach is conservative.

This trade-off will be central to frontier-model competition. A model that is too permissive may create unacceptable risk. A model that is too restrictive may become unusable for the experts most capable of strengthening critical systems. Anthropic faces the same challenge, which is why it separates Fable 5 from the less restricted Mythos 5 configuration.

Sol does not resolve the dilemma. It demonstrates that capability and access policy are becoming inseparable product features. Companies evaluating the model must test not only whether it can perform a task, but whether it will reliably perform that task under the safeguards applied to their account and use case.

Where Sol Still Falls Short

The launch data does not support the claim that GPT-5.6 Sol is the best model at everything. Claude models lead several software-engineering, tool-use and long-context evaluations. Gemini remains highly competitive in multimodality, speed and cost. Grok offers a compelling combination of fast output and live information tools. DeepSeek provides a level of openness and deployment control that Sol cannot match.

Sol’s million-token context also requires careful interpretation. It performs strongly on several retrieval and graph-reasoning tests, but it does not dominate every evaluation at the upper end of the context window. Applications should use retrieval, memory systems and context management rather than assuming they can insert a million tokens and receive perfect reasoning.

Ultra mode introduces another limitation: cost predictability. Parallel agents can complete difficult work faster, but they can also multiply token consumption. A loosely defined task may produce several expensive investigations that do not improve the final answer. Enterprises will need routing policies that determine when multi-agent reasoning is justified.

The model remains capable of hallucination. Tool use can reduce unsupported claims by allowing the system to consult external data, but tools also create new failure modes. The agent may choose the wrong source, misread a result, apply an incorrect transformation or take an action based on a flawed assumption.

Human oversight remains essential in finance, medicine, law, cybersecurity and critical infrastructure. Sol can reduce the amount of supervision required for routine stages of a workflow. It cannot eliminate accountability.

Who Should Use GPT-5.6 Sol?

Sol is best suited to tasks in which failure is costly, the workflow spans several tools and the output has enough economic value to justify premium inference. Complex software engineering, investment research, security analysis, scientific workflows, legal document review, strategic planning and executive-level artifact creation are natural fits.

It is less compelling for high-volume classification, simple summarization, routine customer support or basic content generation. Terra, Luna, Gemini Flash, DeepSeek Flash or other lower-cost models may deliver better economics for those workloads.

The strongest production architecture will often use more than one model. A low-cost model can classify requests, extract data and handle routine interactions. Sol can be called when the task requires deeper reasoning, long-context synthesis, computer use or multi-agent investigation. A specialized model can then validate code, calculations or domain-specific conclusions.

This routing approach reflects the broader direction of AI infrastructure. Companies are unlikely to choose one model for every task. They will build portfolios in which models compete for work based on capability, latency, cost, privacy and risk.

Sol is designed to become the premium escalation layer in that portfolio. Its success will depend on whether it can repeatedly justify the escalation.

A Model Built for the Post-Chatbot Era

GPT-5.6 Sol arrives at a moment when the AI industry is changing its definition of progress. Larger benchmark scores still matter, but they no longer tell the whole story. The decisive questions are whether a model can complete a real workflow, how much supervision it needs, how quickly it can recover from mistakes and what the successful outcome costs.

Sol is OpenAI’s strongest answer to those questions so far. Its combination of reasoning controls, programmatic tool use, large context, computer interaction, artifact generation and optional multi-agent execution makes it more than a conventional language model. It is an attempt to package intelligence as an adaptable operational system.

Claude Fable 5 may remain stronger for certain long-running coding and analytical tasks. Gemini may offer a better balance of speed, price and ecosystem integration. Grok may be more attractive for real-time information and rapid engineering workflows. DeepSeek may be the strategic choice for organizations that prioritize openness, sovereignty and self-hosting.

Sol’s advantage is breadth combined with efficiency. It can reason deeply without always reasoning expensively. It can operate tools without requiring every step to be manually scripted. It can produce polished work rather than stopping at a plausible draft. When the problem becomes unusually difficult, it can escalate from one agent to several.

That does not make GPT-5.6 Sol a universal winner. It makes it a strong candidate for the role that may become most valuable in enterprise AI: the model called when ordinary automation reaches its limit.

The long-term significance of Sol will therefore not be measured by how many users prefer its conversational style. It will be measured by how much difficult work organizations are willing to entrust to it—and how often the model can return with the job genuinely finished.

Anthropic is keeping its most powerful generally available model within reach of paying subscribers for another week. The company has extended included access to Claude Fable 5 through July 19, while also maintaining a temporary 50% increase in Claude Code’s weekly usage limits. For developers and other intensive Claude users, the announcement translates into more room for ambitious projects—but the two benefits come with important limits that are easy to misunderstand.

The extension applies automatically to eligible subscribers. There is no promotional code to enter and no separate trial to activate. Users can continue selecting Fable 5 from supported Claude interfaces, while Claude Code users receive the higher weekly allowance as part of their existing plan.

What Anthropic has not done is make Fable 5 unlimited or permanently bundle it into every subscription. The model can consume only part of a subscriber’s included weekly allowance, and users who cross that threshold must either move to another model or begin paying through usage credits.

Two Different 50% Figures

Anthropic’s announcement contains two separate benefits involving the number 50%, which may create confusion.

The first concerns Fable 5. Eligible subscribers may use the model for up to 50% of their normal weekly usage allowance without an additional metered charge. This does not mean Anthropic is giving users an extra 50% of Fable capacity on top of their subscription. Instead, Fable 5 can consume as much as half of the weekly allowance the account already has.

Once that Fable-specific threshold is reached, the rest of the subscriber’s included weekly capacity remains available for other Claude models. A user could switch to Sonnet 5 or Opus 4.8 and continue working within the remaining allowance. Users who want to stay on Fable 5 can enable usage credits, which move further activity onto consumption-based billing.

The second 50% figure applies specifically to Claude Code. Anthropic is temporarily keeping Claude Code’s weekly usage limits at 1.5 times their standard level. This is additional weekly capacity for the coding product, not a rule limiting Claude Code to half of anything.

In practical terms, a developer who normally receives a certain weekly Claude Code allocation now receives 50% more during the promotion. The account’s shorter five-hour usage window does not receive the same boost, however. A user can still run into a five-hour limit during a particularly intensive session even when substantial weekly capacity remains.

Who Receives the Extended Fable Access

Anthropic describes the offer as covering all paid plans, but its support materials provide a more precise definition. Included promotional Fable 5 usage is available to Claude Pro and Max subscribers, Team customers and eligible premium seats on seat-based Enterprise plans.

Enterprise administrators should pay particular attention to their seat configuration. Standard Enterprise seats have not historically received the same included Fable allowance as premium seats. Those organizations may still make the model available through usage credits, depending on the controls and billing settings established by their administrators.

Consumption-based Enterprise customers and API developers are in a different position. Their access is already metered rather than governed by the consumer-style promotional allowance. The July 19 extension is primarily meaningful for subscription customers who would otherwise have to pay separately to continue using Fable 5.

Free Claude accounts are not included.

The Claude Code limit increase covers eligible Pro, Max, Team and seat-based Enterprise users. Because Claude’s limits can differ by plan and seat type, the most reliable indicator is the usage section inside the account rather than an assumed number of prompts or coding hours.

Why Fable 5 Matters

Fable 5 sits above Anthropic’s Opus line in the company’s capability hierarchy. It shares its underlying model with Claude Mythos 5, a more restricted version intended for approved cybersecurity and research partners, but Fable adds extensive safeguards designed for general deployment.

Anthropic positions Fable 5 as its strongest widely released option for long-running agents, difficult software engineering, complex analytical work, visual reasoning and scientific research. Its advantage is intended to become more visible as tasks grow longer and require the model to maintain a plan across many steps.

That distinction matters in Claude Code. Many coding assistants can generate a function, explain an error or make a small edit. Fable 5 is aimed at work closer to codebase-wide migrations, sustained debugging, architectural changes, autonomous tool use and projects requiring repeated verification.

The model also uses adaptive thinking, meaning it determines how much internal computation to devote to a request. Users can influence that behavior through effort settings, but Fable is designed to reason rather than simply return the fastest possible response.

This capability comes at a cost. Fable 5 can consume subscription limits faster than less expensive models, particularly during long conversations, large repository scans and high-effort agentic sessions. The fact that users may allocate half of their weekly allowance to Fable does not guarantee half a week of continuous use. Actual consumption depends on context size, task complexity, model effort, tool calls and the amount of existing conversation history that must be processed again.

What Happens When the Fable Limit Is Reached

Users approaching the Fable-specific cap should expect Claude to indicate that the model’s included allowance is nearly exhausted. At that point, there are two main paths.

The cost-conscious option is to switch models. Sonnet 5 is Anthropic’s default model on several plans and is designed to offer a more efficient balance of speed and capability. Opus 4.8 remains suitable for complex coding and enterprise work while generally costing less to operate than Fable.

The alternative is to continue with Fable through usage credits. Credits are separate from the subscription fee and are billed according to metered model consumption. Fable 5’s standard pricing is $10 per million input tokens and $50 per million output tokens, compared with $5 and $25 respectively for Opus 4.8.

That difference can become significant when a project includes a large repository, lengthy conversation history or repeated autonomous tool use. Users enabling credits should establish a monthly spending cap rather than relying on manual monitoring alone. Claude’s usage settings allow subscribers to review consumption, set alerts and limit additional spending.

Users are warned before included usage transitions to credits. Anthropic does not silently convert ordinary subscription usage into unrestricted metered billing without the relevant credit configuration and confirmation.

What the Claude Code Increase Changes

The temporary weekly increase is particularly valuable for developers who use Claude Code for sustained work rather than occasional questions. The extra capacity can support more repository exploration, parallel subagents, testing cycles, code reviews and longer implementation sessions before the weekly ceiling becomes the constraint.

It does not remove every form of throttling. Claude Code usage is governed by both short-term and weekly limits. The five-hour allowance controls how intensely an account can use the service over a concentrated period, while the weekly allowance controls cumulative activity over the account’s assigned cycle.

Only the weekly side receives the temporary 50% increase. Developers who encounter the five-hour limit must still wait for that window to reset, reduce the intensity of their workflow or continue through usage credits where available.

Weekly limits also reset according to a fixed schedule assigned to each account. The July 19 deadline does not necessarily coincide with an individual user’s weekly reset. The promotion increases the allowance available during eligible cycles, but unused capacity should not be expected to carry over after the offer ends.

Inside Claude Code, the /usage command can show remaining capacity and the next reset time. The usage dashboard in Claude’s account settings provides the broader picture across supported Claude products.

The Extension Follows an Unusual Launch

Fable 5’s route to general availability has been less straightforward than a typical model rollout.

Anthropic initially launched Fable 5 on June 9. Three days later, the company suspended access after the United States government imposed export controls on Fable 5 and Mythos 5. According to Anthropic, the immediate nature of the restrictions and the difficulty of verifying users’ nationality in real time led it to remove access globally.

The controls were subsequently lifted, and Anthropic restored Fable 5 on July 1 with updated cybersecurity safeguards. The company initially included the model on eligible subscriptions through July 7. It later extended that window to July 12 and has now moved the deadline again to July 19.

That sequence helps explain why access is still being presented as a temporary promotion rather than a permanent entitlement. Anthropic has said demand for Fable is difficult to predict and that it ultimately wants to restore the model as a standard component of subscription plans when capacity permits.

Each extension gives the company more data about real-world demand, compute consumption, safety interventions and the willingness of users to pay for Fable once included access ends.

Safeguards May Cause Automatic Model Switching

Users testing Fable 5 should also expect occasional model switching that has nothing to do with rate limits.

Fable operates with safety classifiers covering areas including offensive cybersecurity, some biology and chemistry requests, and attempts to extract the model’s reasoning or capabilities. When a request triggers one of these systems, Claude may route the task to Opus 4.8 instead of allowing Fable to answer.

Claude should notify the user when this happens. Anthropic says most Fable sessions do not trigger a fallback, although legitimate security, debugging or scientific work may be more likely to encounter one.

The distinction matters because switching to Opus is not necessarily evidence that the Fable allowance has been depleted. It may instead reflect the model’s safety routing. Developers working in dual-use fields should therefore pay attention to the message shown in the interface rather than assuming every model change is caused by consumption.

Fable 5 also carries a 30-day data-retention requirement for covered traffic and is not available under zero-data-retention arrangements. That condition is most consequential for enterprise and API customers handling sensitive workloads, but it reinforces the need to check organizational policy before moving regulated or confidential projects onto the model.

How Users Should Use the Extra Week

The extension is best treated as an evaluation window for demanding work, not as an invitation to route every prompt through the most expensive model.

Fable 5 is likely to deliver the greatest value on tasks where a stronger model can reduce the number of failed attempts, coordinate a long sequence of actions or maintain coherence across a complicated project. Architectural planning, difficult debugging, large migrations, financial analysis, visual reconstruction and research synthesis are better candidates than routine editing or simple code generation.

Sonnet 5 remains the more efficient choice for everyday work. Opus 4.8 provides a middle ground when a task requires greater reasoning depth but does not justify Fable’s higher consumption.

Developers should also consider starting fresh sessions when moving to unrelated tasks. Long conversation histories increase the amount of context the model must repeatedly process. Monitoring effort settings, limiting unnecessary repository context and assigning clear completion criteria can help stretch the promotional allowance.

The same discipline applies to Claude Code’s larger weekly limit. Additional capacity creates the most value when used for well-scoped autonomous work with tests and verification, rather than open-ended sessions that repeatedly inspect the same material.

What Comes After July 19

Unless Anthropic announces another extension, the current promotion ends at 11:59:59 p.m. Pacific Time on July 19. After that deadline, Fable 5 is expected to require usage credits for subscription customers rather than drawing from the included promotional allowance.

Claude Code’s weekly limits are also expected to return to their standard levels. The permanent increases Anthropic previously made to five-hour limits remain separate from this temporary weekly promotion.

A further extension is possible, given that Anthropic has already moved the Fable deadline more than once. Users should not plan business-critical workflows around that possibility, however. The safer assumption is that metered Fable billing and ordinary Claude Code weekly limits will resume after the announced cutoff.

For now, paying subscribers have another week to determine whether Fable 5 produces enough additional value to justify its higher consumption. The most important expectation is not unlimited access, but controlled access: half of the existing weekly allowance for Fable, 50% more weekly room in Claude Code and a clear return to metered economics once the promotion closes.