Anthropic shipped Claude Fable 5.1 yesterday, September 1, 2026, and I spent the evening running it on my own stack instead of reading about it. The short version: same price as Fable 5, cache reads cut by 75%, fewer false refusals, and a model that stays coherent across a three-hour session of unglamorous infrastructure work.
Fable 5 was the best model I had used and also the one people complained about most. I wrote about the launch-day rebuild it did on a legacy CRM→ and about the export-control pause and its return→. A point release is Anthropic's answer to the complaint list. So the question for 5.1 is narrow: did they fix the things that hurt, and did the price move.
What changed in Claude Fable 5.1
Per Anthropic's announcement, the per-token price is unchanged at $10 per million input tokens and $50 per million output. The change is cache reads: $0.25 per million, down 75%. Anthropic estimates that makes typical workloads around 25% cheaper and highly agentic ones up to around 45% cheaper, because agent loops resend the same context on every turn.
The second lever is effort. Anthropic says Fable 5.1 at low or medium effort matches or beats Fable 5 at a much lower cost. For an operator that is the bigger deal. Most of what an agent does in a day is routine, and paying `xhigh` prices for routine work was the main reason Fable 5 bills ran away.
Safeguards moved too. Cybersecurity safeguards block 60% fewer false positives per session, biology safeguards fire 85% less often on benign biology and medical questions, and Fable 5.1 may now identify software vulnerabilities for defensive work, though not write exploits. Mythos 5.1 is the same model with fewer safeguards, limited for now to a set of US organizations: the Life Sciences Verification Program is built around it, while the Cyber Verification Program still runs Opus- and Sonnet-class models and is slated to add Mythos-class models "in the near future."
The behavioral notes are the ones I care about. Anthropic claims 5.1 avoids shortcuts that produce poorer work, does better root-cause analysis, and holds coherence over long multi-step tasks. Their flagship example is Millennium: a crash that hit roughly one run in a million and had gone unexplained for years. Fable 5.1 disassembled a vendor library, matched it against a core dump, and traced the bug into that library.
Availability is immediate: `claude-fable-5-1` on the API, claude.ai, Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry.
Fable 5.1 benchmarks against Fable 5, Opus 5 and GPT-5.6 Sol
Anthropic's own table. Vendor-reported, so treat the gaps as directional until independent leaderboards catch up.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| --- | --- | --- | --- | --- |
| Terminal-Bench 4.0 (agentic coding) | 55.8% | 42.0% | 52.3% | 37.3% |
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| AutomationBench (business workflows) | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
| GDPval-AA v2 (knowledge work) | 1853 | 1723 | 1824 | 1711 |
| Humanity's Last Exam (no tools) | 60.9% | 57.8% | 56.6% | – |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% | – |
Two things stand out. Terminal-Bench-Science, Terminal-Bench 4.0 and AutomationBench are the long-horizon agent tests, and they moved the most: the science benchmark more than doubled, from 24.7% to 52.6%, and the other two each gained about 14 points. And Opus 5, which many people fell back to when Fable 5 got expensive or stubborn, now sits clearly behind on the same tasks. Cognition said they are moving Devin's Opus 5 traffic to Fable 5.1 on launch day, and cited the cache pricing as the reason a Fable-class model is now affordable for them.
One caveat from the other side of the fence: on OpenAI's Agents' Last Exam, GPT-5.6 Sol→ beat Fable 5 by 13.1 points, and at medium reasoning still led by 11.4 points at roughly a quarter of the estimated cost. Anthropic did not publish a 5.1 number for that benchmark. Sol still costs half as much per token, $5 in and $30 out, so the price argument for 5.1 rests on cache reads and low effort, not list price.
The complaint list 5.1 was built for
The Fable 5 backlash had a consistent shape. A five-hour usage cap on the web app. Safety classifiers that quietly handed your conversation to Opus 4.8. Heavy token consumption, with one widely shared report of $110 burned in a day. Laziness deep into long tasks, verbosity, slowdowns mid-run, and time-to-first-token past a minute on hard prompts. "A Ferrari with a 30 mph limiter" was the line that stuck. Opus 5 got its own complaints, and Thariq, who works on Claude Code at Anthropic, agreed on X that it is "a really spiky model" and that fixing it is "a huge priority for us."
During the grayscale rollout in the last week of August, testers who got routed to the new checkpoint reported better long-chain reasoning, no more forgetting preconditions halfway through a task, and better tool use and step planning. Some testers estimated 10 to 20% better results on legal work. The one place it got stricter: it refuses to generate images that involve copyrighted material.
That matches the official framing. Fewer false refusals, less laziness, lower cost per completed task. The usage cap on the consumer app is the item I have not seen addressed.
What builders need to change
If you call the API directly, Fable 5.1 has a few sharp edges that Fable 5 did not, per Anthropic's migration guide:
None of these bit me in Claude Code, which handles them for you. They will bite anyone with a hand-rolled agent loop that forces tool calls.
An evening running Fable 5.1 on my own stack
Benchmarks measure the model on someone else's problems. Mine, last night, were three hours of the kind of work that fills an operator's evening: a broken tool, two database migrations, two deploys, and a widget that had never been connected to Spotify. Fable 5.1 in Claude Code did all of it, and I reviewed as it went.
The broken tool. My site's MCP server has a `search_console` tool that had been returning Unauthorized for weeks. The cause was two layers deep. The process that served my tool calls was not the one defined in the repo's MCP config; Claude Desktop launched a second one through a shell script that never received the right key. And production still ran an older auth rule, because a 48-file working tree with the new one had never been committed. The process table showed an unexpanded placeholder where a key should be, and the model refused to guess. It found the process that served my calls by sampling TCP connections to my domain during a tool call, then attached Node's inspector to the process the repo config had spawned and confirmed the placeholder was real. The fix was one Node flag in two files.
The migrations. The Supabase MCP connector returned permission denied for DDL. The Supabase CLI choked on my `.env.local`. A Vercel env pull came back with `[SENSITIVE]` placeholders instead of connection strings, and the model briefly wrote those placeholders into my env file before catching it and reverting. A linked-project run from a symlinked directory could not find the project ref. Four dead ends, and it still got the migrations through. I have watched earlier models loop on the first failed route; this one kept changing approach.
The deploy. Seventy files, two migrations, three new Vercel variables, a build, a push, a production deploy, and verification of the deployment state and the affected endpoints, including the negative case: the legacy key now correctly gets a 401 on the owner-only route.
The widget. A Spotify now-playing footer widget existed in the tree, but the OAuth step had never been run; the redirect URI was not even registered in the Spotify dashboard. Fable 5.1 walked me through the OAuth setup, tested the route locally, and then ran an 18-agent adversarial review of the feature: three reviewers with different lenses, every finding handed to a separate agent whose job was to refute it. 15 findings went in, 5 survived. The survivors were real: artist text at 2.04:1 contrast against a 4.5:1 requirement, physical padding that broke the pill on my Arabic route, missing fetch timeouts, an uncaught upstream error that would surface as an uncached 500, and a refresh token that would expire in six months and vanish silently. It fixed all five, measured the contrast and the RTL padding in a browser, and deployed. I left the refresh token out of production on purpose, so the widget stays hidden there. When I asked for it, it logged the state so the next session knows what is live. Nine minutes for the review, about 1.65 million tokens across the 18 agents.
Now the receipts against it. Twice that evening it printed a secret into tool output, once my Spotify client secret and once a short-lived access token, through a shell expansion mistake while checking whether variables were set. It caught both, told me, and told me to rotate the secret. It also spent close to an hour on the question of which process served my MCP calls before landing on a two-line fix. That hour was the model refusing to accept evidence that contradicted itself, which is the behavior I want, but it is still an hour.
Which model to compare it with
Compare it with Fable 5 first. A point release is a repair, so the test is whether the repairs landed. That evening: no laziness late in the session, and it held the thread across three hours, two deploys and an 18-agent review. Refusals I did not count, and one evening is not a sample. The cost cut I did not measure either; it rests on Anthropic's cache price and the effort setting, and I tracked neither.
Compare it with GPT-5.6 Sol second. Sol is half the list price and still owns the long-horizon agentic score on OpenAI's own benchmark. If your workload is high-volume and forgiving, Sol remains the cheaper seat. If it is the kind of work where one wrong shortcut costs you a day, Fable 5.1 is the model I would hand it to.
Do not compare it with Gemini yet. Google's flagship is still Gemini 3.1 Pro and has no fresh numbers. And keep an eye on OpenAI's Astra, rumored for the first half of September. If it lands, this post gets an update.
Disclosure, in the spirit of this blog: Fable 5.1 drafted this post inside the same session it describes, from the day's sources and its own tool log. I reviewed every claim and every number.
