On September 1, 2026, Anthropic released Claude Fable 5.1 at the same sticker price as its predecessor. Cache reads got 75% cheaper, and on the agentic scientific task benchmark the model more than doubled Fable 5's score. Neither fact tells the whole story.
A ".1" label normally points to a minor update. Claude Fable 5.1 isn't one, at least not in any category that matters for demanding AI work. Calling it a generational leap without qualification would be just as wrong, though, because the benchmark table holds both kinds of numbers. To judge this model fairly, you have to separate the truly large gains from the narrow ones, and see where the cost story is more complicated than the press release suggests.
What follows is what the data actually shows: the findings that matter more than the headline number, and the claims that need independent verification before you change anything in production.
What it is and where it fits
Claude Fable 5.1 is Anthropic's most capable generally available model, released on September 1, 2026 as the successor to Claude Fable 5. It's also the first version where Anthropic states the Mythos relationship openly at launch. Claude Mythos 5.1 uses the same underlying model weights as Fable 5.1 under different safeguard configurations, and it is open only to vetted organizations through Anthropic's Cyber Verification Program and Life Sciences Verification Program.
The specifications carry over from Fable 5: a 1-million-token context window, 128K maximum output, and adaptive thinking that is always on, with depth set through the effort parameter (default: high). The knowledge cutoff moves to June 2026, and the API identifier is claude-fable-5-1. The model is available through the Claude Platform and API, Claude Code, Amazon Bedrock, Cursor, OpenRouter, and GitHub Copilot.
Quick spec summary
Context: 1M tokens | Max output: 128K tokens | Thinking: Adaptive, always on
Input: $10/M tokens | Output: $50/M tokens | Cache reads: $0.25/M tokens
Cache writes: $12.50/M (5-min), $20/M (1-hour) | Batch API: $5/$25 per M tokens
Knowledge cutoff: June 2026 | API ID: claude-fable-5-1
The pricing change that actually matters
The sticker price stayed put: input and output tokens still cost $10 and $50 per million, just as they did for Fable 5. What changed is the price of cache reads, which fell from $1.00 to $0.25 per million tokens, a 75% reduction, or a 4x cut in cache-read costs.
Whether that shows up on your bill depends entirely on how you use the model. Anthropic estimates the cache-read cut lowers costs by about 25% for typical workloads and about 45% for highly agentic ones, where cached context makes up a large share of total token consumption. Agents keep returning to the same system instructions, codebase context, tool definitions, and conversation history. Once that context is written to cache, it can be read many times at $0.25 instead of $1.00 per million tokens, and the savings add up quickly at scale.
The savings claim comes with a significant caveat, so read it before you update your cost model. Artificial Analysis, which worked with Anthropic on pre-release testing, found that at maximum effort Fable 5.1 costs roughly 20% more per Intelligence Index task than Fable 5, because it produces about 1.7 times as many output tokens. At max effort, Fable 5.1 runs about $3.76 per Intelligence Index task, against $2.34 for Opus 5, which scores only three points lower. The cache-read cut is real, but how it nets out on your bill depends on the ratio of output tokens to cached input tokens in your own workload.
The benchmark numbers: where the gap is real
At launch, Anthropic published a nine-benchmark table comparing Fable 5.1 with Fable 5, Opus 5, and GPT-5.6 Sol. Every result was run by Anthropic with production safeguards switched on. That last detail matters, and the caveats section below explains why.
Terminal-Bench-Science 0.1 is the number in every headline, and its size earns the attention. The benchmark measures whether a model can autonomously carry out real scientific research tasks in a terminal environment. Fable 5.1 reaches 52.6%, compared with 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. The standard error is ±3.5 to 4.5 points per model, so some model-to-model comparisons carry uncertainty larger than the gaps between them. Fable 5.1's lead of 23 to 30 points, however, sits well outside that noise band. A model scoring 52% on autonomous scientific research in a terminal is doing something categorically different from one scoring 24%.
Terminal-Bench 4.0, which tests multi-step agentic coding and debugging, puts Fable 5.1 at 55.8%, against 42.0% for Fable 5 and 52.3% for Opus 5. The gain over Fable 5 is large. The lead over Opus 5 is 3.5 points, which is real but modest enough that the cost comparison between the two belongs in any decision about which to deploy.
AutomationBench has Fable 5.1 at 31.4%, against 17.1% for Fable 5 and 19.6% for GPT-5.6 Sol, a big step up on business workflow automation. GDPval-AA v2, an Elo-style measure of professional knowledge work quality, gives Fable 5.1 an Elo of 1853, against 1824 for Opus 5, 1723 for Fable 5, and 1711 for GPT-5.6 Sol. The 130-point jump over Fable 5 is bigger than the 101-point jump Opus 5 made over Fable 5, which means the earlier Opus-tier deficit on knowledge work has reversed within a single release.
"Fable 5.1 is not uniformly smarter; it is much stronger where a task runs long and uses tools, and only marginally stronger on work its predecessor already did competently."
Emergent.sh, Claude Fable 5.1 Benchmarks: Scores and What They Mean, September 2026
The benchmark numbers: where the gap is narrow
CursorBench 3.2.0, which tests coding inside an IDE workflow, gives Fable 5.1 73.4%, against 70.5% for Fable 5 and 70.0% for Opus 5. That's a 3.4-point lead over Opus 5, on a model that costs exactly twice as much per token. Fable 5.1 wins, but the win doesn't justify the price premium for everyday coding, unless your own evals show your tasks hitting a ceiling that Opus 5 can't clear.
HLE (Humanity's Last Exam), which tests multidisciplinary reasoning, puts Fable 5.1 at 60.9% without tools and 65.0% with tools, against 56.6% and 63.6% for Opus 5. The gap is genuine, at 4.3 points without tools and 1.4 with them, but it isn't decisive on a benchmark where sampling variance and prompt sensitivity play a real part.
OSWorld 2.0, which scores computer-use tasks in a real operating system environment, reports Fable 5.1 at 77.9% partial / 41.7% strict, against 75.4% / 39.6% for Opus 5. These figures use the benchmark authors' August 2026 task release, which differs from earlier OSWorld releases, so they can't be compared with previously published OSWorld results. For that reason, Anthropic published no GPT-5.6 Sol scores for this benchmark. Both Fable models scored zero on tasks where safeguards intervened.
The honest caveats about these numbers
Anthropic ran every benchmark in its launch table itself. That's not unusual for an AI model launch, but it's the right context for reading the numbers. Anthropic's benchmark methodology has generally held up against independent replication, yet launch tables are still the vendor's best presentation of a model's abilities. The sensible response is to run your own evals on your actual tasks before making Fable 5.1 a production default.
The safeguard zeros deserve a closer look. Both Fable 5.1 and Fable 5 scored zero on OSWorld 2.0 tasks where production safeguards intervened, and Fable 5 scored zero on some AutomationBench tasks for the same reason. Anthropic notes that the published scores "likely reduce the performance of Fable 5.1 and Fable 5 on these benchmarks." In other words, the raw capability ceiling is higher than the scores show, and the safeguards cost real performance points.
Mythos 5.1, which runs the same weights with less restrictive safeguards for vetted organizations, scores 60.9% on Terminal-Bench 4.0 against Fable 5.1's 55.8%. That 5-point gap is the measurable cost of the safety configuration. It's also the cleanest evidence in the launch table that the published Fable 5.1 numbers understate what the underlying model can do.
Independent analysts confirmed the direction of all the gains. Artificial Analysis ranks Fable 5.1 first on its Intelligence Index with a score of 66, ahead of Opus 5 at 63 and GPT-5.6 Sol at 61. ARC-AGI-1 puts it at 97.5%, and Vals AI ranks it first on the Vals Index. The direction of the benchmarks is well supported. The size of individual results, especially those involving computer use and narrower coding leads, would benefit from independent replication on your specific workload.
Real-world signal at launch
Anthropic shared one reported result from Millennium, an investment firm with early access: Fable 5.1 traced a rare software crash to a bug inside an external vendor library, after the problem had defied explanation for four to five years. It's a single data point, and a vendor-curated one, but it's specific enough to work as a signal. The model's strength on long-running diagnostic tasks with deep context is a practical capability, not just a benchmark category.
Anthropic also reported results on three protein design targets from Adaptyv Bio's competitions, citing a near-50% hit rate on protein binders. Whether those results hold up under independent testing should become clear in the coming months.
Breaking changes developers need to know now
If you run Fable 5 in production, two changes need a code review before you switch the model identifier.
First, forced tool use behaves differently. Setting tool_choice to any or to a named tool now returns a 400 error instead of processing the request. Any agent or workflow that relies on forced tool selection to guarantee a specific tool fires must be updated before the model swap.
Second, thinking block compatibility has changed. Earlier Claude models can't read Fable 5.1's extended thinking blocks, although Fable 5.1 can read those of earlier models. If you're building multi-model workflows where thinking blocks pass between models, you'll need to handle that asymmetry. Editing earlier conversation turns also invalidates thinking blocks, and this rule is strictly enforced for API accounts created on or after August 31, 2026.
Safety filter improvements worth noting
One underreported part of the Fable 5.1 launch is a recalibration of the safety filters. In Fable 5 and Opus 5, the cybersecurity and biology filters were sensitive enough to trigger on well-intentioned requests at a frustrating rate. Fable 5.1's cybersecurity filters produce 60% fewer false positives than their predecessors, and biology-related filters fire 85% less often on harmless questions about basic biology and medicine. For developers building security tooling, research assistants, or anything in adjacent domains, that's a meaningful quality-of-life improvement the benchmark table doesn't capture.
Who should switch and who should stay on Opus 5
Anthropic's own guidance is unusually direct here. Fable 5.1 is positioned for demanding reasoning and long-horizon agentic work, while Opus 5 is the recommended starting point for most workloads. Fable 5.1 is the upgrade to reach for when Opus 5 at higher effort still falls short on your specific evaluations.
These are the cases where Fable 5.1 earns its 2x premium over Opus 5:
Long-running agent loops with heavy cached context. The 75% cache-read reduction narrows the cost gap substantially. An agent that reads a 500K-token codebase hundreds of times in one session pays $0.25 per million tokens read instead of $1.00. At scale, that materially shrinks the effective cost premium.
Autonomous scientific research or analysis in terminal environments. The Terminal-Bench-Science jump is no rounding difference. A capability gap of 52.6% versus 29.0% is real and large enough to justify an evaluation.
Complex production incident diagnosis. Long context, cross-file reasoning, and multi-step tool use. The Millennium example is vendor-curated, but the underlying task profile matches what the benchmarks show to be Fable 5.1's strong suit.
Business workflow automation at scale. AutomationBench shows nearly double Fable 5's score. If workflow automation is a primary use case and your own evaluation confirms the gain, the cost justification is there.
Opus 5 is probably the better call for everyday coding assistance (the CursorBench lead is 3.4 points at twice the price), for general writing and knowledge work (a 29-point GDPval Elo gap isn't enough to override the price difference in most applications), and for any task where your current Opus 5 evaluation results are already acceptable.
"Fable 5.1 is the generally available model for demanding reasoning and long-horizon agentic work. Anthropic recommends starting with Opus 5 for most workloads, and reaching for Fable 5.1 when your evals on Opus 5 at higher effort still fall short."
Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1, September 1, 2026
The Mythos 5.1 tier and what it signals
Over the long term, the most structurally significant part of this launch is Anthropic making the Mythos/Fable relationship public. Same weights, two safety configurations, two access tiers. Fable 5.1 is what the production safeguards allow, and Mythos 5.1 is what the underlying model can do when safety filtering is calibrated for vetted, specialized use cases rather than general deployment. The 5-point Terminal-Bench 4.0 difference between them puts a number on one cost of the general-availability safety configuration.
That gap will likely narrow as Anthropic keeps calibrating its safeguards. The 60% drop in cybersecurity false positives and the 85% drop in biology filter triggers in Fable 5.1, relative to earlier versions, show the direction of travel. The goal is a model that is both safe to deploy generally and not penalized on every technical task that touches a sensitive domain. Fable 5.1 is closer to that goal than Fable 5 was, and the benchmarks reflect it.
The short version: switch to Fable 5.1 if your workloads involve long-running agents with heavy cached context, autonomous agentic tasks in scientific or engineering domains, or complex diagnostic reasoning over large codebases, and run your evals first. Stay on Opus 5 if everyday coding or general knowledge work is your main use case. The cache-read cut is the most significant practical change in this release, and it pays off most for teams whose compute costs come mainly from re-reading cached context rather than generating new output.
Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.
I do freelance scraping and data automation projects. Feel free to contact me on LinkedIn or email at hi@kawsarlog.com
Top comments (0)