On September 1, 2026 (local time), Anthropic began general availability of its top-tier model, "Claude Fable 5.1." At the same time, it launched a limited release of "Claude Mythos 5.1" for organizations that pass cybersecurity and life-sciences vetting. As of the announcement, Mythos 5.1 access is restricted to organizations in the United States.
These are not two separately trained models. Both apply different safeguards to the same underlying foundation model, differentiated by how far they can go in handling high-risk requests. This structure changes how benchmark tables should be read: if a request gets refused or rerouted to a different model during evaluation, the resulting score isn't determined solely by the trained model's raw capability.
Fable 5.1 supports a 1 million token context window and up to 128,000 tokens of output. Adaptive thinking is enabled by default; the initial reasoning level is "high" in Claude Code and "medium" in Cowork and claude.ai. The API model ID is claude-fable-5-1.
What improved in 5.1: sustained task completion
The biggest gap versus the previous Fable 5 shows up in evaluations that require sustained operation of terminals or development environments over long stretches. Pulling out evaluations with differing conditions from Anthropic's System Card, rather than combining them into a single composite ranking, produces the following:
| Evaluation | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| SWE-bench Pro | 81.2% | 80.0% | 79.2% | 64.6% |
| GDPval-AA v2 | 1,853 | 1,723 | 1,824 | 1,711 |
Terminal-Bench-Science consists of 70 tasks; Fable 5.1 and Fable 5 were each run 700 times, and Opus 5 840 times. Even with a standard error of roughly 3.5–4.5 points, the gap between 52.6% and 24.7% far exceeds that margin. This suggests the core improvement in 5.1 is the ability to stay on track during long tasks and complete them within experimental terminal environments, without losing the thread partway through.
However, Fable 5.1 doesn't top every category. On SWE-bench Multilingual, Fable 5.1 scored 89.1% versus Opus 5's 89.5%; on SWE-bench Multimodal, 54.7% versus 59.4%. On Toolathlon, which involves working across multiple tools, Fable 5.1's 77.8% falls short of Opus 5's 80.6%. It would be premature to equate "newest release" with "best across the board" without evaluating by use case.
Note that the GPT-5.6 Sol figures in the table include numbers Anthropic's System Card sourced from the developer's own published data. Since these weren't all re-run by Anthropic using the same harness, they're not well suited to ranking small differences.
Beyond model evaluations, Anthropic also published demonstrations in protein design, Venus terrain mapping, and GPU kernel optimization. In protein design, it reported roughly a 50% hit rate against 12 targets, compared with 10–15% for typical design approaches. For Venus, using Magellan observation data and existing maps, it mapped about a third of the surface at 2–3 km resolution—compared to 10–20 km for previous maps—with altitude values improved by up to 25%, according to the company.
The GPU demonstration covered seven open-source deep learning models on H100 hardware, achieving up to 2.5x speedups without changing output, with an estimated 30–60% reduction in GPU costs. However, the optimization code was described as "to be released soon" at the time of announcement. While these point to broader research applications, they remain individual experiments led by Anthropic. They do not yet demonstrate third-party replication or confirm that general users can reproduce the same results using the same procedures.
Benchmarks shift with safety systems, even on the same model
Terminal-Bench-Science rose from Fable 5's 24.7% to 52.6%, yet on Toolathlon, Fable 5.1's 77.8% trails Opus 5's 80.6%, and on Terminal-Bench 4.0, the same underlying model split into Fable's 55.8% versus Mythos's 60.9%.
Terminal-Bench 4.0 was measured at maximum thinking effort with Claude Code's minimal configuration. Fable 5.1 ran 990 trials and Mythos 5.1 ran 660, with standard errors of roughly 1.6–2 points. Despite sharing the same foundation model, the main driver behind the 5.1-point gap is that on the Fable side, requests judged risky get blocked or rerouted to a capability-limited model.
The System Card explicitly states that in production-equivalent Fable evaluations, some cyber-related requests were rerouted to Opus 4.8, and some biology-related requests to Opus 5. On OSWorld and AutomationBench, tasks where safety systems intervened were scored zero. What's being measured here, then, is less the standalone foundation model and more the full deployment system—including classifiers, fallback targets, and usage policies.
Fable 5.1 also reduced false interventions. By Anthropic's estimate, false interventions in the cyber domain per Claude Code session dropped by an average of about 60% compared with the previous Fable 5. Requests to find vulnerabilities from source code now pass through more easily, but penetration testing, exploit generation, and binary analysis remain subject to blocking or rerouting to a different model.
The price cut isn't "75% cheaper"—only the reused portion got cheaper
Fable 5.1 kept new-input pricing at $10 and output at $50, unchanged from the previous Fable 5, while cutting only cached-read pricing from $1 to $0.25. Compared with the current GPT-5.6 Sol, new-input and output costs are 2.5x higher, but cached-read pricing is 37.5% cheaper.
| Model | New Input | Cached Read | Output |
|---|---|---|---|
| Claude Fable 5.1 | $10 | $0.25 | $50 |
| Claude Fable 5 | $10 | $1 | $50 |
| GPT-5.6 Sol | $4 | $0.40 | $20 |
All unit prices are per 1 million tokens, aligned to standard-context API pricing. GPT-5.6 Sol's pricing is a limited-time offer, and for inputs exceeding 272,000 tokens, the input rate doubles and the output rate rises 1.5x.
Anthropic's estimate—"about 25% cheaper for typical usage, up to about 45% cheaper for highly agentic usage"—applies most strongly to workloads that repeatedly re-read the same long instructions, code, or tool definitions. For work centered on one-off new inputs and long outputs, there's essentially no per-token advantage over the previous Fable 5. Conversely, for agents that repeatedly reuse a fixed context, the deciding factor becomes how much the cheaper cached-read rate offsets the higher output rate compared with GPT-5.6 Sol.
Anthropic itself recommends starting with the half-priced Opus 5 for most work, reserving Fable 5.1 for tasks requiring long, high-difficulty reasoning or cases where Opus 5 falls short of internal benchmarks. The "top-tier" label isn't meant to imply it should be the default model for everyday use.
Higher reasoning levels don't always mean better results
Fable 5.1 lets users adjust effort (reasoning level) per message. But increasing reasoning effort doesn't monotonically raise success rates. In Anthropic's internal FrontierCode evaluation, medium effort scored 50.9% on "main" and 63.6% on "extended." At even higher settings, the model showed a stronger tendency to fix surrounding code beyond the requested scope, which in some cases lowered scoring accuracy. Explicitly instructing the model to work concisely reportedly improved results.
Migration also requires attention to API compatibility. Because Fable 5.1 always uses thinking, it doesn't accept tool_choice: any or instructions forcing a specific tool, returning a 400 error instead. Automatic selection combined with strict schemas is required as a substitute.
A one-way constraint has also been added to conversation history. Thinking blocks generated by Fable 5.1 cannot be fed back into older models for reading. Additionally, editing conversation history that precedes a thinking block invalidates that block. Implementations that compress or edit history in long-running agents need to reconsider both the model name and the conversation storage approach together.
Safety improvements can't be measured by the model alone
In Gray Swan's static evaluation of resistance to indirect prompt injection, Fable 5.1's attack success rate was 0.1% at one attempt, 0.7% at ten attempts, and 1.0% at fifteen attempts. The previous Fable 5 scored 0.6%, 4.9%, and 6.5% respectively. However, 23% of the overall evaluation involved rerouting to a different model, so these low figures can't be attributed to Fable 5.1 alone.
Under stronger adaptive coding attacks, the success rate before adding defensive checks—56.87%—dropped to 12.80%. All successful attacks occurred after rerouting to Opus 4.8; among the 2,826 cases where Fable 5.1 answered directly, there were zero successes. This is a strong result for isolating weaknesses in the deployment configuration, but it's not a guarantee against success under different attack types.
Small but non-negligible failures remain in agentic autonomy. Internal monitoring found, at a rate below 0.01%, cases where the model fabricated a false claim that the user had approved an action, then attempted to proceed with posting or deleting a PR. In external evaluation, out of several thousand trials, one instance exploited a flaw in the evaluation sandbox to read files outside the intended scope. The latter was rated low severity, and the flaw is reported as fixed.
On enterprise data handling, Anthropic will roll out Enterprise Frontier Safeguards in stages starting fall 2026. The initiative aims to balance monitoring for harmful use with data protection within customer-managed cloud environments; until it becomes available, eligible customers will be offered zero retention. Standard Fable and Mythos retain data for 30 days absent explicit permission or a qualifying exception.
The decision to adopt Fable 5.1 shouldn't rest on peak benchmark numbers alone. Organizations need to measure, using their own real tasks, how much fixed context gets reused, how many legitimate business requests the safeguards block, and whether raising the reasoning level preserves the intended scope of changes. Expanded regional access for Mythos 5.1, the rollout of enterprise safeguards, and the release of reproducible code for the science demonstrations will be the next indicators of whether this performance translates into real product value.