Anthropic pushed out two new flagship models on September 1, 2026: Claude Fable 5.1 and Claude Mythos 5.1. The headline number is a 75% cut to cache-read pricing, but the more consequential story sits underneath it, in the compute buildout that makes a model like this possible in the first place. Anthropic says its run-rate revenue has now surpassed $30 billion, up from roughly $9 billion at the end of 2025, and the company is racing to secure gigawatts of Trainium, TPU, and Nvidia GPU capacity to keep up with a customer base that more than doubled its million-dollar accounts in under two months.
Fable 5.1 and Mythos 5.1 arrive twelve weeks after their predecessors, Fable 5 and Mythos 5, which launched June 9 and were briefly pulled from the market after a government export-control directive. That backstory matters here, because Anthropic built the safeguard changes in Fable 5.1 directly in response to what happened in June. This piece breaks down what actually changed in the models, what it costs, what hardware sits behind it, and what it means for the wider AI infrastructure race involving Amazon, Google, Microsoft, and OpenAI.
What Anthropic Actually Shipped on September 1
According to Anthropic’s own launch page, Claude Fable 5.1 and Claude Mythos 5.1 are the same underlying model shipped with different levels of safeguards. Fable 5.1 is generally available today through the Claude API, Claude.ai, Claude Code, and Claude Cowork, and it runs on Amazon Web Services, Google Cloud, and Microsoft Azure. Mythos 5.1 is restricted to vetted participants in Anthropic’s trusted-access programs, aimed specifically at cybersecurity defenders and life-sciences researchers.
Anthropic frames the pair as its most capable models yet for coding, knowledge work, and long-running research tasks, and says their scientific research abilities offer an early look at how AI systems could start contributing directly to lab work rather than just supporting it. Developers can start building against the model immediately using the identifier claude-fable-5-1 in the API. Thurrott’s coverage of the release notes that the launch lands as enterprise customers keep pushing Anthropic on price, data retention, and false-positive rates from earlier safeguards, three areas the company addresses directly in this update.
Claude Fable 5.1 vs Claude Mythos 5.1: One Model, Two Safeguard Tiers
The distinction between Fable 5.1 and Mythos 5.1 isn’t architectural. Both are built from identical weights. What separates them is how tightly Anthropic’s safety classifiers intervene during a session. Fable 5.1 keeps standard safeguards in place and will now decline to help build exploits for software vulnerabilities it identifies, even though it can help find them. Mythos 5.1 loosens those same classifiers for a narrow set of approved users working inside two formal channels: the Cyber Verification Program and the Life Sciences Verification Program.
Anthropic enrolled its first Life Sciences Verification Program participants in coordination with the US government and plans to widen access to the broader research community over time. On the cyber side, the Cyber Verification Program already covers Opus- and Sonnet-class models with reduced restrictions for defensive security work, and Anthropic says Mythos-class access is coming to that program soon. Claude Security, the company’s product for scanning codebases and suggesting patches, now runs on Mythos 5.1 as its underlying engine.
The Benchmark Numbers, Explained
Anthropic published head-to-head scores against Fable 5, its own Opus 5, and OpenAI’s GPT-5.6 Sol across six benchmarks. Claude Fable 5.1 posted its largest gain on Terminal-Bench-Science 0.1, a test of agentic scientific research tasks, more than doubling the score of Fable 5. It also widened its lead on AutomationBench, a benchmark built around simulated business workflows.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 (research) | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 (agentic coding) | 55.8% | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 (knowledge work) | 1,853 | 1,723 | 1,824 | 1,711 |
| Humanity’s Last Exam (with tools) | 65.0% | 63.8% | 63.6% | — |
| AutomationBench (business workflows) | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 (agentic coding) | 73.4% | 70.5% | 70.0% | 67.2% |
Two caveats apply to these figures. Anthropic tested Fable 5.1 with production safeguards active, and on tasks where those safeguards intervened, the model scored zero rather than being excluded from the average. And Terminal-Bench-Science carries a fairly wide standard error, on the order of 3.5 to 4.5 points per model, so the gap between Fable 5.1 and its rivals on that specific test should be read as directionally large rather than precise to a decimal.
Anthropic also ran Terminal-Bench 4.0 with Mythos 5.1’s more permissive safeguards switched on, where it posted 60.9%, five points above Fable 5.1 on the identical underlying model. That gap illustrates just how much of the score difference in agentic coding benchmarks currently comes down to how often a safety classifier interrupts a task rather than any change in raw capability.
Why the 75% Cache Cut Changes the Math
The single most concrete number in this launch is pricing. Fable 5.1 keeps the same headline rate as Fable 5, at $10 per million input tokens and $50 per million output tokens. What changed is the cost of cache reads, the charge applied when the model reuses context it has already processed rather than reprocessing it from scratch. That rate drops from $1.00 to $0.25 per million tokens, a 75% reduction.
| Pricing metric | Fable 5 | Fable 5.1 | Change |
|---|---|---|---|
| Input tokens (per million) | $10.00 | $10.00 | No change |
| Output tokens (per million) | $50.00 | $50.00 | No change |
| Cache reads (per million) | $1.00 | $0.25 | -75% |
| Indexed cost, typical workload | 100 | ~75 | -25% |
| Indexed cost, highly agentic workload | 100 | ~55 | -45% |
Typical Workloads vs Highly Agentic Runs
Anthropic measured the indexed cost figures above across four weeks of actual August 2026 usage on Claude Enterprise, Claude Code, and the API, at default effort settings. A typical workload, spanning general enterprise use, ends up costing about 25% less overall. A highly agentic workload, the context-heavy, tool-heavy pattern where cache reads dominate total spend, drops by closer to 45%. For teams running long unattended coding sessions or multistep research agents, where the model repeatedly re-reads the same growing context window, that second figure is the one that actually shows up on the invoice.
Both models default to a 1 million token context window, with a maximum output of 128,000 tokens per response. Fable 5.1 defaults to high effort inside Claude Code and medium effort in Claude Cowork and on Claude.ai, a setting users can now adjust mid-conversation rather than restarting a session to change how hard the model works on a given step.
The Hardware Stack Behind Every Claude Query
Anthropic has been explicit that it does not train or serve Claude on a single chip family. In its announcement of an expanded Google and Broadcom partnership, the company states that it trains and runs Claude across AWS Trainium, Google TPUs, and Nvidia GPUs, matching workloads to whichever chip suits them best. That diversification is now the backbone of the entire Fable 5.1 and Mythos 5.1 rollout, since Claude is the only frontier model currently available to customers on all three major cloud platforms at once, through Amazon Bedrock, Google’s Vertex AI, and Microsoft Foundry.
| Cloud partner | Chip families | Committed capacity | Reported commitment |
|---|---|---|---|
| Amazon Web Services | Trainium2, Trainium3, Trainium4, Graviton | Up to 5GW; ~1GW online by end of 2026 | $100B+ over 10 years |
| Google Cloud + Broadcom | TPU v8, TPU v9, custom Broadcom TPUs | 3.5GW starting 2027, on top of 1GW+ already live in 2026 | Tens of billions of dollars |
| Microsoft Azure | Nvidia GPUs (via Foundry) | Part of Anthropic’s three-cloud availability strategy | Not separately disclosed |
The AWS relationship, detailed in earlier coverage of Amazon’s Trainium buildout and corroborated by The New Stack’s reporting on the AWS investment, secures Anthropic up to 5 gigawatts of Trainium and Graviton capacity, with nearly a gigawatt of Trainium2 and Trainium3 due online by the end of this year. Amazon remains Anthropic’s primary training partner through the joint Project Rainier cluster. The newer Google and Broadcom arrangement, which followed Anthropic’s earlier TPU compute agreement with Google and was also detailed by Silicon Republic, adds 3.5 gigawatts of custom Broadcom-built TPU capacity starting in 2027, stacked on top of more than a gigawatt already running in 2026.
A $30 Billion Run Rate and the Compute Arms Race
Anthropic CFO Krishna Rao described the Google and Broadcom expansion as a continuation of a disciplined approach to scaling infrastructure, framing it as the company’s most significant compute commitment to date, made necessary by the pace of customer growth rather than a one-off bet. The numbers back that up: Anthropic’s run-rate revenue has crossed $30 billion, up from about $9 billion at the close of 2025, and the number of customers spending over $1 million annually has passed 1,000, doubling in under two months from the 500-plus the company reported at its February Series G raise.
Most of the new compute capacity is being sited inside the United States, which Anthropic ties to a broader November 2025 pledge to put $50 billion into domestic computing infrastructure. Every dollar of that spending is ultimately what makes a 75% cache-price cut possible without Anthropic taking a margin hit, since cheaper, more diverse compute lowers the underlying cost of serving inference at scale.
Mythos 5.1 in the Lab: Venus Maps, Protein Binders, and GPU Kernels
The most unusual part of this launch is what Anthropic showed in scientific research rather than software engineering. Given access to open-source protein design and folding tools, Mythos 5.1 designed binder proteins that external validators found to be up to ten times more potent, in terms of binding affinity, than the best entries in Adaptyv Bio’s public protein design competitions, on three separate target proteins. Its overall hit rate across twelve targets reached close to 50%, against a typical industry hit rate of 10% to 15%.
Separately, Anthropic says Fable 5.1 trained a neural network that produced a new high-resolution elevation map covering roughly a third of the surface of Venus, built from three-decade-old radar data captured by NASA’s Magellan mission. The new map resolves surface features down to two to three kilometers, compared with 10 to 20 kilometers previously, and Anthropic is releasing it under a Creative Commons license ahead of the upcoming NASA VERITAS and ESA EnVision missions to the planet.
On the hardware-optimization side, Mythos 5.1 wrote custom GPU kernels for seven open-source genomics and protein-modeling tools, speeding them up by as much as 2.5 times on Nvidia H100 hardware while producing identical outputs. For labs running these models across thousands of variants, Anthropic estimates the optimization cuts GPU costs on genome-wide analyses by 30% to 60%, work the company says would typically require weeks from a dedicated performance-engineering team. Anthropic also used this launch to preview the Model Hardware Standard, a specification that lets Claude agents directly and safely operate physical laboratory equipment, currently in research preview with a first group of scientific labs and manufacturers.
Tighter Safeguards, Fewer False Positives
Anthropic rebuilt its safety classifiers between Fable 5 and Fable 5.1 specifically to cut down on benign requests getting flagged as dangerous. The company says its updated biology safeguards now fire 85% less often on ordinary medical and elementary biology questions compared with the classifiers that shipped with Fable 5, though anything touching life-sciences research and development still routes to Anthropic’s Opus models. On the cybersecurity side, the newest safeguards produce 60% fewer false positives, and Claude Code users should see roughly 60% fewer safeguard interventions per session than they did under Fable 5’s older classifiers.
That precision cuts both ways. Fable 5.1 is now allowed to help identify software vulnerabilities, defensive work that was previously blocked outright, but tasks like penetration testing, exploit generation, and binary-based vulnerability scanning still get redirected to Opus-class models regardless of the requester. Anthropic also added anti-distillation protections, closing off a documented technique that let outside parties manually edit Claude’s prior conversation context while preserving its internal reasoning trace, a method used to extract model capabilities at industrial scale through networks of fake accounts.
The Shadow of June: Export Controls, Suspension, and Recovery
None of these safeguard changes happened in a vacuum. Fable 5 and Mythos 5 originally launched June 9, and just three days later, on June 12, the US government issued an export control directive after Amazon researchers reported a way to prompt Fable 5 into identifying software vulnerabilities, in one case generating code that demonstrated how to exploit one. Anthropic pulled access to both models for all users while it addressed the issue.
The government lifted the export controls on June 30, with Fable 5 restored globally on July 1 and Mythos 5 restored to a set of US organizations following government approval on June 26. Anthropic used the intervening weeks to work with Amazon, Google, Microsoft, and other partners in its Project Glasswing program on a shared industry framework for triaging the severity of AI jailbreaks, and to deepen pre-release testing cooperation with the US government. Fable 5.1’s more precise cybersecurity classifiers are the direct product of that six-week sprint, and it’s part of why Anthropic is emphasizing testing depth this time around, including external evaluations from two outside organizations plus automated red-teaming from Gray Swan.
Enterprise Frontier Safeguards: The Zero-Retention Pitch
Alongside the model release, Anthropic is rolling out Enterprise Frontier Safeguards, a system built with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, retail, and government, in partnership with AWS, Google Cloud, and Microsoft Azure. EFS stores customer data on infrastructure the customer controls rather than Anthropic’s own systems, with any required human review handled by the customer, aiming to match the privacy of a zero-data-retention agreement while still catching adversarial misuse. It’s rolling out in phases starting this fall across Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Google’s Agent Platform, and Microsoft Foundry. Until it’s generally available, eligible enterprise customers can already run Fable 5.1 under a standard zero-data-retention policy.
Anthropic is also complying with the EU AI Act’s Code of Practice on Transparency of AI-Generated Content, which it signed alongside 190 other companies in July 2026. Every model Anthropic has released since August 2, 2026 embeds an invisible watermark indicating the likelihood that Claude generated a given piece of text. The company is rolling out a detection API in private preview to regulators, law enforcement, media organizations, and researchers who are legally entitled to check for it.
Competitive Landscape: Where This Leaves OpenAI and Google
Anthropic’s own benchmark table puts GPT-5.6 Sol behind Fable 5.1 on every metric it published, including a wide gap on Terminal-Bench-Science and AutomationBench. Those numbers come from Anthropic’s own testing, so they should be read as a vendor’s framing of the comparison rather than an independent audit, but the direction of the release is unmistakable: Anthropic is competing on cost and defensive-security capability at the same time it competes on raw benchmark scores. That combination is harder for rivals to match quickly, since a 75% cut to cache pricing changes the economics of long-context, agentic workloads in a way that a marginally higher benchmark score does not.
The bigger competitive story may be structural rather than model-specific. By running on Trainium, TPUs, and Nvidia GPUs across all three major clouds simultaneously, Anthropic has positioned itself as the only frontier lab that isn’t dependent on a single chip supplier or a single cloud partner for capacity. OpenAI’s relationship with Microsoft Azure and Google’s internal TPU-first strategy each carry more single-vendor concentration than Anthropic’s current setup, at least based on what all three companies have disclosed publicly through 2026.
Market Impact and Enterprise Buying Decisions
For enterprise buyers, the practical effect of this release is a pricing conversation more than a capability one. Anthropic’s own workload modeling shows the biggest savings land on exactly the use cases enterprises are scaling fastest: agentic coding, document-heavy research, and tool-calling workflows that keep large amounts of context resident across many turns. A 45% cost reduction on that category of workload is a large enough number to shift procurement conversations, particularly for firms already running Opus-class models for tasks where Fable-class pricing was previously too close to make switching worthwhile.
That dynamic also explains why Anthropic is pushing so hard on infrastructure diversification. Every gigawatt of Trainium, TPU, or GPU capacity it locks in reduces the odds that a shortage on any single chip family constrains how aggressively it can cut prices for customers. With more than 1,000 customers now spending above $1 million annually, and that number reportedly doubling in under two months, Anthropic has real incentive to keep the unit economics of inference moving in its customers’ favor rather than just its own.
Historical Context: Twelve Weeks From Fable 5 to Fable 5.1
The jump from Fable 5 to Fable 5.1 is one of the fastest point-release turnarounds Anthropic has shipped, and it stands out against a backdrop of a company that has moved from roughly $9 billion in annualized revenue to more than $30 billion inside a single year. That growth trajectory is part of what has fueled speculation about a future Anthropic IPO, even as the company continues raising private capital and signing multi-year infrastructure deals instead. The pattern across 2026 has been consistent: ship a model, secure more compute, expand safeguards after real-world testing exposes a gap, then ship the next point release with those lessons built in. Fable 5.1 and Mythos 5.1 fit that pattern almost exactly, arriving with pricing cuts funded by new compute capacity and safeguards rebuilt in direct response to the June export-control episode.
It’s also worth comparing this release to Anthropic’s other current model line. Readers who followed the earlier comparison between Opus 5, Mythos 5, and Gemini 3.1 Pro will recognize the pattern here: Anthropic keeps its most permissive, highest-risk model access gated behind vetted programs while pushing capability and price improvements out to the broadly available tier first. The prior Fable 5 vs. Opus 5 vs. GPT-5.6 Sol comparison is a useful baseline for judging exactly how much ground Fable 5.1 closed against Anthropic’s own higher-tier model in twelve weeks.
Getting Started With Claude Fable 5.1
Developers already on the Claude API can switch to the new model by updating the model identifier in an existing integration. A minimal request looks like this:
curl https://api.anthropic.com/v1/messages
-H "x-api-key: $ANTHROPIC_API_KEY"
-H "anthropic-version: 2023-06-01"
-H "content-type: application/json"
-d '{
"model": "claude-fable-5-1",
"max_tokens": 1024,
"messages": [
{"role": "user", "content": "Summarize the pricing changes in Fable 5.1."}
]
}'Teams already invested in the Claude ecosystem covered in our own hardware and AI chips coverage should expect the cache-pricing change to apply automatically to any workload billed by token, without needing to opt in, since Anthropic applied it as a blanket rate cut rather than a separate pricing tier.
What Comes Next: Five Predictions
- Expect Anthropic to widen Enterprise Frontier Safeguards availability faster than its fall timeline suggests, given how many regulated-industry customers helped build it.
- The Cyber Verification Program will likely extend Mythos-class access within the next quarter, given Anthropic’s own statement that this is already planned.
- Rival labs, particularly OpenAI, will face pressure to cut their own cache or context-reuse pricing rather than compete purely on benchmark scores.
- Anthropic’s compute partners will announce additional gigawatt-scale capacity commitments before the 3.5GW Broadcom TPU tranche comes fully online in 2027.
- The Model Hardware Standard research preview will expand beyond its first cohort of labs and manufacturers if early safety testing holds up, given Anthropic’s pattern of gradually widening access to new capabilities.
None of these are certainties, and Anthropic has not committed to specific dates for most of them. But they follow directly from the pattern the company has set across 2026: expand compute, cut prices where the new capacity allows it, and widen restricted-access programs once safety testing catches up with capability.
Frequently Asked Questions
What is the difference between Claude Fable 5.1 and Claude Mythos 5.1?
They are the same underlying model. Fable 5.1 ships with standard safeguards for general use, while Mythos 5.1 loosens those safeguards for vetted cybersecurity and life-sciences professionals inside Anthropic’s trusted-access programs.
How much cheaper is Claude Fable 5.1 than Fable 5?
Input and output token pricing stays at $10 and $50 per million tokens. Cache-read pricing drops 75%, from $1.00 to $0.25 per million tokens, which cuts typical workload costs by around 25% and highly agentic workload costs by around 45%, according to Anthropic’s own measurements.
What is the context window for Fable 5.1 and Mythos 5.1?
Both default to a 1 million token context window with a maximum output of 128,000 tokens per response.
Can I access Claude Mythos 5.1 directly?
Only through Anthropic’s Cyber Verification Program or Life Sciences Verification Program. Mythos 5.1 is not available on the general Claude API the way Fable 5.1 is.
Why were Fable 5 and Mythos 5 pulled from the market in June 2026?
The US government issued an export control directive on June 12 after researchers found a way to prompt Fable 5 into identifying and, in one case, demonstrating an exploit for a software vulnerability. Anthropic suspended access to both models until the controls were lifted on June 30.
What hardware does Anthropic use to run Claude?
Anthropic trains and serves Claude across AWS Trainium chips, Google TPUs, and Nvidia GPUs, and is the only frontier AI lab currently available to customers on all three major cloud platforms: Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.
Is Claude Fable 5.1 better than GPT-5.6 Sol?
On the six benchmarks Anthropic published comparing the two models, Fable 5.1 scored higher than GPT-5.6 Sol across every category, including agentic coding and business-workflow automation. These figures come from Anthropic’s own testing rather than an independent third party.
What is Enterprise Frontier Safeguards?
It’s a new privacy system Anthropic built with more than 100 enterprise customers that stores data on infrastructure the customer controls, aiming to match zero-data-retention privacy while still detecting misuse. It rolls out in phases starting this fall.