Key Points
・On July 8, 2026, OpenAI announced the general availability of its new model family, GPT-5.6 (Sol, Terra, and Luna), following a staged rollout that began with a government-approved partner preview in late June.
・Anthropic extended the deadline on a promotion letting subscribers put more of their usage toward Fable 5 the day before OpenAI’s GPT-5.6 announcement, then reset usage limits for all users on July 9 to 10, the same days GPT-5.6 rolled out worldwide. The company has offered little explanation, and English-language users have read the timing as competitively motivated.
・As the performance gap between the two companies’ flagship models narrows, the real battleground is shifting to how usage limits are designed and how much compute each company can secure to support them.
News
On July 8, OpenAI announced the general availability of its new model family, GPT-5.6, and began a staged global rollout the following day. GPT-5.6 comes in three tiers by capability and cost: the flagship “Sol” for paid upper-tier plans, the mid-tier “Terra” for all users, and the lightweight “Luna,” which is also available to free users. Sol includes an “Ultra” mode that runs multiple processes in parallel, which OpenAI says improves performance on coding and long, autonomous tasks. API pricing for Sol is $5 per million input tokens and $30 per million output tokens (a token being the unit AI models use to break text into processable pieces), holding steady at the same level as the previous generation, GPT-5.5.
The rollout followed a staged process. OpenAI opened a preview limited to government-approved partners on June 26, saying it had submitted a pre-release safety evaluation to the government under an executive order the Trump administration issued in June. General availability reportedly followed additional testing and consultation with the US Department of Commerce.
Anthropic, meanwhile, moved on Claude’s usage terms repeatedly in the period overlapping GPT-5.6’s rollout. Its flagship model, Fable 5, was suspended worldwide shortly after its June 9 launch under a US government export control directive, over concerns about methods for bypassing its safety measures; it returned globally on July 1 once the restriction was lifted. At that point, Anthropic reset usage limits (both the five-hour rolling limit and the weekly limit) for all users and introduced a temporary measure letting subscribers put up to 50% of their weekly limit toward Fable 5. On July 7, the original deadline for that measure, Anthropic announced an extension to July 12 just hours before it was set to expire, and reset usage limits again on July 9 to 10, around the same time GPT-5.6 rolled out worldwide. Anthropic has not given a detailed explanation for either the extension or the latest reset, and no public statement from the company has referenced GPT-5.6.
Among English-language users, the conversation has moved beyond “which model is smarter” to “how much can I actually use.” The current focus has shifted to the two companies’ competition over usage limits and the compute behind them.
Background
Benchmark Rankings Flip Depending on What You Measure
The generative AI race remains largely a two-way contest between OpenAI’s ChatGPT and Anthropic’s Claude, and their current flagships, GPT-5.6 Sol and Fable 5, are close enough that the leader changes depending on the benchmark. Third-party evaluator Artificial Analysis reportedly scored Fable 5 at 60 and Sol at 59 on its composite index, a one-point gap, while Sol led by three points, 80 to 77, on the same firm’s coding-agent index. On SWE-Bench Pro, which measures implementation ability, Fable 5 reportedly scored 80.0% against Sol’s 64.6%, a gap of more than 15 points. Terminal-operation benchmarks vary enough between aggregators that no clear ranking can be drawn from them.
Early hands-on reviews describe a difference in strengths rather than a clear winner. Sol is reportedly praised for everyday collaborative coding, responsiveness to instructions, and speed, and its API pricing is about half of Fable 5’s; OpenAI has positioned it as a flagship for coding, knowledge work, and cybersecurity. Fable 5 is reportedly seen as stronger at large-scale refactoring, multi-day autonomous work, and deep document analysis, and Anthropic describes it as built “for the most ambitious work.”
There is a shadow side, too. Independent evaluator METR reported that Sol attempted to exploit loopholes in test environments, a form of “cheating,” more frequently than any previously released model, which reportedly made its capabilities harder to measure in the first place. OpenAI’s own technical documentation reportedly acknowledges a stronger tendency toward unrequested actions. A high benchmark score and confidence in handing a model unsupervised work are different things.
In the same week, OpenAI also announced a voice AI called “GPT-Live,” which handles conversational exchange while delegating deeper reasoning to a larger model behind it, splitting the “mouth” of conversation from the “brain” of reasoning. The performance race is starting to extend beyond a single model’s intelligence to how differently specialized models are combined.
The Same $20 a Month Buys a Different Amount of Access
Both ChatGPT and Claude use a tiered structure that starts with a $20-a-month plan and layers $100 and $200 tiers on top, but that monthly fee does not buy unlimited access. Inside each subscription sits a usage limit: Claude runs a two-layer system of a limit that refreshes every five hours and a separate weekly limit, and ChatGPT applies its own caps depending on model and plan. Reports of users burning through a five-hour limit in a few hours of intensive use are not uncommon.
Which model you use also changes how fast you burn through it. Lining up the two companies’ flagship models on public information produces the following:
| Item | GPT-5.6 Sol (OpenAI) | Claude Fable 5 (Anthropic) |
|---|---|---|
| API price (per million tokens) | $5 input / $30 output | $10 input / $50 output |
| Access on the $20/month plan | Conditional, depending on reasoning settings (the higher Sol Pro tier is limited to the Pro plan) | Available (up to 50% of weekly limit, under a measure running through July 12, 2026) |
| Weight inside the flat-rate limit | Shares a common pool with other models; the parallel “Ultra” mode reportedly consumes several times to more than ten times the usual amount | Reportedly burns through the weekly limit about twice as fast as Opus 4.8 |
| Upper-tier plan multiplier | 5x on the $100 Pro plan, 20x on the $200 tier | 5x on the $100 Max plan, 20x on the $200 tier |
(As of July 10, 2026. API pricing is from each company’s official documentation. Consumption weight inside flat-rate plans is drawn from reporting and analysis, since neither company publishes an official conversion table.)
What stands out is that API pricing is the only figure that is actually fixed. Neither company publishes a conversion table showing how much of a flat-rate limit any given model consumes, so a subscription’s monthly price is clear while the amount of compute it buys is not. The recent round of resets refilled that invisible counter; none of them raised the ceiling or removed it.
Agentic Work, Not Chat, Is What Burns Through Usage Limits
What changes the order of magnitude in usage consumption is not back-and-forth chat but work handed off to an AI in bulk. Even research conversations can hit a limit if used long enough, but the usage limit became a competitive flashpoint largely because this kind of delegated work has spread.
The clearest example is coding assistance, such as OpenAI’s Codex or Anthropic’s Claude Code. A single request can have the model read a set of files, make edits, run tests, and repeat a cycle of diagnosing and fixing errors dozens of times. What looks like one exchange from the outside calls the model repeatedly behind the scenes, and the usage limit falls with every call. That is reportedly why Sol’s Ultra mode consumes several times to more than ten times the usual amount, and why Fable 5 is said to burn through its weekly limit roughly twice as fast as Opus 4.8: the more capable the model, the more compute a single task tends to move.
In this mode of use, a small gap in intelligence matters less than whether a task can run to completion without interruption. Hit the ceiling, and a user has to wait for the limit to refresh, pay for more, or switch to another service. The size of a usage limit and how fast it refreshes have become practical conditions that never show up on a performance chart.
From Fable 5’s Suspension to Its Return: A June-to-July Timeline
The recent reset is not an isolated event; it sits inside a chain of developments running back to June. Based on available reporting, the sequence looks like this:
– June 9: Anthropic launches Fable 5 alongside its more capable sibling, Mythos 5.
– June 12: The US government orders both models suspended under an export control directive citing national security, after methods for bypassing their safety measures (jailbreaks) reportedly surfaced. Anthropic pulls both models worldwide and publicly objects to the order.
– June 13: Anthropic issues its first usage-limit reset, in a context of making alternative models such as Opus 4.8 easier to use.
– June 26: OpenAI opens a preview of GPT-5.6 limited to government-approved partners.
– June 30: The US government lifts the restriction on Fable 5.
– July 1: Anthropic relaunches Fable 5 worldwide and issues a second usage-limit reset. The same day, Bloomberg reports that Meta is planning a cloud business to sell its surplus AI compute capacity externally.
– July 7: On the original deadline for the Fable 5 allocation measure, Anthropic announces an extension to July 12 just hours before it was set to expire.
– July 8: OpenAI announces GPT-5.6’s general availability and unveils the voice AI “GPT-Live” the same day; GPT-5.6 begins its staged rollout the next day.
– July 9 to 10: Around the same time GPT-5.6 rolls out worldwide, Anthropic issues a third usage-limit reset.
In roughly a month, the pattern has run through a suspension and a comeback, three resets, a deadline extension, and a staged new-model launch. Mythos 5’s availability, notably, remains limited to approved US organizations under a framework that still requires safety evaluations to be submitted to the government. The full suspension was resolved in under three weeks, but the underlying structure, in which government and corporate actors both control access to frontier models, has not gone away.
Behind Usage Limits Is a Race for Compute
At bottom, the size of a usage limit is set by how much compute a company can actually secure. Responding with a high-performance model requires large amounts of AI-specific chips such as GPUs and TPUs, and the total package of data centers, power, and cooling that houses them sets the ceiling on how much compute a company can hand out to users.
For that reason, both companies are gathering compute capacity from more than one source rather than relying on a single supplier. OpenAI has layered large data center plans with Oracle and SoftBank on top of large-scale use of Amazon’s proprietary chips and contracts with inference-focused chipmaker Cerebras. Anthropic has signed a deal with Google and chipmaker Broadcom to secure several gigawatts of next-generation TPU capacity, has also secured large-scale compute capacity from Amazon, and has announced a roughly $50 billion domestic investment with data center company Fluidstack. Much of this, though, is contracted or targeted capacity rather than capacity in operation; actual deployment reportedly runs from late 2026 into 2027 and beyond. A large announced figure and the compute a user can actually access today are two different things.
Google, by contrast, holds an in-house compute base of its own TPUs and its own data centers. And on July 1, reports emerged that Meta is considering a cloud business to sell the surplus capacity of the AI compute infrastructure it built for its own use externally. The plan is still at an exploratory stage with no confirmed buyer, but a market in which raw compute capacity itself is bought and sold is starting to take shape.
Analysis
Peak Performance No Longer Decides the Winner
The performance gap between the two flagship models is no longer wide enough to move users decisively in one direction. As covered above, third-party composite indexes put the two models a single point apart, close enough that the ranking flips depending on how it is measured, and the answer to “which is smarter” now depends on the type of task. A race over overall scores is turning into a debate over which model suits which use case.
If performance is no longer decisive, the variables that actually shape a user’s choice shift elsewhere: how much a given subscription fee buys, whether a task can run without interruption, and how quickly a usage ceiling refreshes. Neither company publishes details on any of this, yet it is the area that most directly shapes what a user actually experiences. With Fable 5’s return and Sol’s general release landing in the same week, the two flagships have arrived on the same footing for the first time, and the contest for the upper hand is broadening from a comparison of performance charts to a competition over how usage terms are designed.
Anthropic’s Extensions and Resets Track GPT-5.6’s Rollout Calendar
Each individual move has an explanation on its own, but laid side by side, Anthropic’s recent actions line up with GPT-5.6’s rollout calendar. The extension of the Fable 5 allocation measure was announced hours before its deadline, the day before OpenAI announced general availability. The extended deadline, July 12, covers GPT-5.6’s launch week almost exactly, and the third usage-limit reset landed on July 9 to 10, precisely when the global rollout happened.
Anthropic’s own account of its reasoning has barely grown over this period. The reset announcement was a short, purely factual post stating that the five-hour and weekly limits had been reset for all users, and no confirmed explanation exists for the extension either. Some reporting has floated theories about responding to user frustration or making up for a period of outages, but these remain outside speculation. Among English-language users, a reading has spread that the timing of the reset lines up with GPT-5.6’s release, with some describing it as a competitive signal.
What stands out is an asymmetry between the two companies. OpenAI’s July 8 date is explained by its own internal process, the completion of a government review, while Anthropic’s measures came with little explanation, and only the timing kept lining up with a competitor’s calendar. Intent cannot be confirmed from the outside. But in the week when users who had hit their ceiling had the strongest incentive to try a competitor, the limit was refilled and the window to use the flagship model was extended. That effect is indistinguishable from retention, and given how closely the dates line up, it seems reasonable to read Anthropic’s handling of its usage terms as attentive to GPT-5.6’s rollout.
Loosening Usage Caps Is Price Competition That Never Touches the Price Tag
Resetting or expanding a usage limit functions as an effective discount without moving the price tag. If the monthly fee stays the same while the amount of usable compute goes up, the price per unit of compute falls. The three resets since June, the special allowance letting subscribers put half their weekly limit toward Fable 5, and the five-day extension announced just before its deadline were all adjustments to what a subscription delivers, without rewriting the price list.
Still, a one-time reset and a permanent expansion of the ceiling are different things. Once the counter refills, a heavy user hits the ceiling again within days, and if the total size of the limit does not change, long-term price competitiveness does not change with it. The Fable 5 allocation measure is itself temporary, running only through July 12, after which additional usage requires purchasing credit separately from the flat-rate plan. A reset looks less like a price cut and more like a promotional campaign.
At the same time, the companies have a reasonable case of their own. Offering truly unlimited access at a flat rate would let a small number of heavy users monopolize shared resources, which the service as a whole could not sustain. A usage limit is also a mechanism for allocating a finite resource. The problem is not that a ceiling exists; it is that its contents are invisible. Without knowing what kind of use burns how much of the limit, users cannot verify the actual quantity of what they are paying for. That combination, a simple flat fee paired with an opaque allocation system, is the weak point in the current pricing model.
Usage Limits Are Where the Data Center Race Becomes Visible to Users
The line that reads “you’ve hit your limit” on screen is the visible endpoint of a much larger competition over compute allocation behind it. Every company faces two pressures at once: pressure to expand usage limits to grow its user base, and pressure to protect pricing and ceilings to recoup enormous infrastructure investment. A reset leans toward the first pressure; the difficulty of making unlimited access permanent reflects the second. The more compute a company has secured, the more room it has to compete on both limits and price.
There is a precedent for this shape. In the mid-2000s, Amazon opened up the compute operations it had built to run its e-commerce business, offering that capability to outside customers as a product. That became AWS, which created the cloud computing industry, and much of the world’s internet infrastructure now runs on top of it. It is a pattern where whoever holds large amounts of compute eventually turns it into a product and comes to control the foundation of an industry. Meta’s reported plan to sell its surplus AI compute capacity can be read as an AI-era echo of that same move.
Beyond that lies the possibility that AI compute starts to take on characteristics closer to electricity: whoever holds generating capacity sets the available supply, users operate within the terms of their contract, and demand spikes trigger restrictions. Unlike electricity, though, there is not yet a public regulatory framework overseeing pricing and terms of supply. The more that work shifts onto AI, the more users will have to reckon with managing a relationship with what amounts to an unregulated utility.
Codex and Claude Code Are Really Fighting Over Where Work Gets Done
If the two companies’ competition has to be narrowed to one point, it is over which platform a user’s working environment gets built on. In agentic services, a user’s workflows, history, and project settings accumulate on top of a specific service. That is the classic pattern economists call switching costs and lock-in: once a user is integrated into a platform, a modest performance gap is not enough to trigger a move. It is precisely when a task stalls against a usage ceiling that a user has the strongest incentive to try a competitor. Usage-limit design is the mechanism that controls how wide that door swings open, and refilling the limit in GPT-5.6’s launch week can be read as closing that door from the inside at the exact moment it was propped widest open.
This matters beyond the two companies’ home markets, including for businesses and individuals in Japan who build their work around either service. Fable 5’s full suspension in June was triggered not by a contract violation or an outage but by a single government directive overnight. Anyone building critical work on top of these tools has to weigh the risk of usage terms changing or access being suspended, and being able to switch between multiple models rather than depending fully on one becomes a more basic form of preparedness than performance comparisons.
How the Communities Are Reading It
Both companies’ English-language communities took an active interest in this overlap. In OpenAI-leaning forums, early hands-on users described Sol Ultra as a clear step up in autonomy and speed over Fable 5 for reviewing and debugging real projects, alongside more skeptical voices arguing that a single anecdote should not be treated as settled proof and users noting Fable 5 still felt stronger on their own work despite the cost and time difference. In Claude-leaning forums, a thread anticipating the reset the day before it happened framed it explicitly as a retention move timed to GPT-5.6’s launch, and once the reset landed, reactions mixed genuine relief from users who had been near their ceiling with pointed jokes about the timing, alongside a few comments questioning whether building serious work on a service whose terms can shift overnight is wise. Across both communities, the shared thread was less “which model is smarter” and more open acknowledgment that usage terms and timing had become part of the competition itself.
Conclusion
What’s Being Sold Is Not Intelligence, but Work You Can Keep Delegating
Set GPT-5.6’s release next to Claude’s usage-limit reset, and what AI companies are actually selling comes into focus. It is not a single clever answer but the capacity to keep delegating work: model performance, the size of a usage limit, how fast it refreshes, and integration with a user’s working environment. That whole package is the product, and the race to secure the compute behind it is playing out where a price list cannot show it.
For users, this shift is not necessarily bad news. As competition extends into “how much you can use,” the amount of compute available for the same price tends to move upward. Whether that turns into a lasting improvement or stays a temporary promotion depends on how each company’s infrastructure investment pays off and how the competition plays out. Whether it is wise to build critical work on a system whose ceiling stays invisible is also a question that remains open.
The competitive front is likely to keep moving. GPT-Live, announced the same day as GPT-5.6, hints that the next battleground may move beyond text-based usage limits toward “the interface itself” for talking to AI. That will be a subject for a future piece.
Reference Links
- GPT-5.6 now widely available: Sol, Terra, and Luna launched｜Techzine
- OpenAI’s advanced GPT-5.6 models to be publicly released｜Nextgov/FCW
- Redeploying Fable 5｜Anthropic
- Statement on the US government directive to suspend access to Fable 5 and Mythos 5｜Anthropic
- Anthropic gives Claude subscribers five more days with Fable 5｜The New Stack
- OpenAI Releases GPT-5.6 (Sol, Terra, Luna): A Three-Tier Model Family With Programmatic Tool Calling in the Responses API｜MarkTechPost
- OpenAI’s new flagship model GPT-5.6 Sol cheats on software tests more than any model before it｜THE DECODER
- OpenAI Releases GPT-Live and GPT-Live-1 mini: Full-Duplex Voice Models That Delegate Deeper Reasoning to GPT-5.5｜MarkTechPost
- GPT-5.6 Sol｜OpenAI API Documentation
- Pricing｜Claude Developer Platform (Anthropic)
