Anthropic released Claude Fable 5.1 on 1 September 2026 at exactly the same list price as Fable 5: $10 per million input tokens, $50 output. The context window did not move. The model ID barely moved.
That combination makes it easy to file under “minor point release” and move on. It is not minor. The most consequential change sits in a pricing footnote, and three breaking changes will return 400 errors on agent code that worked last week.
In this guide:
- What Claude Fable 5.1 Actually Is
- The Cache Read Cut Is the Real Story
- Three Breaking Changes
- The Effort Tiers Cost 3.2x Apart
- Benchmarks and One That Deserves Scrutiny
- Mythos 5.1: Same Weights, Different Safeguards
- Claude Fable 5.1 vs GPT-6 Astra
- Who Should Switch
- Frequently Asked Questions About Claude Fable 5.1
- Verdict
- Sources and Further Reading
What Claude Fable 5.1 Actually Is
Claude Fable 5.1 is Anthropic’s top-end model, released 1 September 2026 as the general-availability successor to Fable 5.
| Specification | Value |
|---|---|
| API ID | claude-fable-5-1 |
| Context window | 1M tokens |
| Max output | 128K tokens |
| Knowledge cutoff | June 2026 |
| Modalities | Text and image in, text out |
| Thinking | Adaptive, always on |
| Input / output | $10 / $50 per million |
| Cache reads | $0.25 per million |
Availability covers the Claude API, Amazon Bedrock as anthropic.claude-fable-5-1, Google Cloud and Microsoft Azure.
Anthropic positions it for demanding reasoning and long-horizon agentic work rather than as a general default, which is a more honest framing than most launch copy and worth taking at face value.
The Cache Read Cut Is the Real Story
Here is what the unchanged Claude Fable 5.1 headline price hides.
Cache reads fell from $1.00 to $0.25 per million tokens, a 75% cut. Cache writes sit at $12.50 per million for the five-minute tier and $20 for the one-hour tier. The Batch API halves input and output to $5 and $25.
Anthropic’s own estimate is that typical workloads cost roughly 25% less than Fable 5, and highly agentic work up to 45% less, entirely from that one line.
Why it matters depends on your workload shape. A single-turn chat request reads little from cache and sees almost no benefit. An agent loop re-reads the same context on every step, so cache reads can dominate the bill entirely. If you run Claude Code, a long-running agent, or anything that keeps a large system prompt warm across many turns, this release cuts your costs by a quarter to nearly half without changing a line of code.
The corollary is blunt: Fable 5 is now the expensive twin for cache-heavy work. There is no reason to stay on it.
This is the same lesson as our LLM API cost optimization guide. Rate cards predict spending badly, and the line item that actually dominates your bill is rarely the one in the headline.
Three Breaking Changes
Buried in the Claude Fable 5.1 migration guide, and they will break working code rather than degrade it.
Forced tool use now returns an error. Code that required the model to call a specific tool gets a 400 rather than a response. This is the one most likely to hit existing agent implementations.
Thinking block handling changed. Adaptive thinking is always on in Claude Fable 5.1, and how thinking blocks are returned and must be passed back differs from Fable 5.
Migration is not a drop-in ID swap. The combination above means changing claude-fable-5 to claude-fable-5-1 in a config file is not sufficient for agent code, even though it is fine for simple completions.
If you run anything agentic in production, read Anthropic’s migration guide before switching. A 400 error on a forced tool call in a scheduled pipeline fails silently in the worst way: the job runs, the error is logged, and nobody notices until the output is wrong.
The security-adjacent thinking in our prompt injection and AI agent security guide applies to failure modes too. Agents fail in ways that look like working.
The Effort Tiers Cost 3.2x Apart
This is the Claude Fable 5.1 number that should shape your configuration and it appears in almost no coverage.
Artificial Analysis measures five effort settings, and the spread is dramatic:
| Effort | Intelligence Index | Output speed | Cost per task |
|---|---|---|---|
| Low | 47 | 53 t/s | $2.37 |
| Medium | 49 | 54 t/s | $2.98 |
| High | 51 | 55 t/s | $3.91 |
| xHigh | 53 | 60 t/s | $5.98 |
| Max | 53 | 65 t/s | $7.63 |
Read the two ends together. Going from low to max costs 3.2 times more per task and buys six index points, from 47 to 53. And xHigh scores identically to Max at 53 while costing $1.65 less.
Two practical conclusions.
Max effort is rarely worth it. It matches xHigh on intelligence and costs 28% more. If you are running max by default, you are paying a premium for output speed rather than quality.
Low effort is underrated. At $2.37 per task it is the cheapest configuration and still scores 47, within six points of the ceiling. For workloads where the answer is checkable or the task is routine, that is the sensible default.
The right pattern is routing by task difficulty rather than picking one tier, which is the same argument this site makes about routing between models entirely.
Benchmarks and One That Deserves Scrutiny
Anthropic’s self-reported Claude Fable 5.1 figures, safeguards enabled:
| Benchmark | Claude Fable 5.1 |
|---|---|
| Terminal-Bench-Science 0.1 | 52.6 (Fable 5: 24.7) |
| Terminal-Bench 4.0 | 55.8 |
| Humanity’s Last Exam | 60.9 / 65.0 |
| OSWorld 2.0 | 41.7 strict / 77.9 partial |
| CursorBench 3.2.0 | 73.4 |
| GDPval-AA v2 | 1853 Elo |
Independent aggregation puts it at 84.6 out of 100 and ranked first of 232 models, with Agentic as its strongest category at number one. Its weakest eligible category is Reasoning, at number two.
The figure worth pausing on is Terminal-Bench-Science more than doubling, from 24.7 to 52.6 in a point release. That is an enormous jump for an increment, and large single-benchmark leaps deserve scrutiny rather than applause. It may reflect genuine capability gains in scientific terminal work, or a change in how the benchmark is run, or a model newly tuned toward a specific evaluation. Anthropic has not broken the improvement down.
Treat it as interesting rather than settled until someone independent reproduces it.
Mythos 5.1: Same Weights, Different Safeguards
An unusual governance arrangement around Claude Fable 5.1 worth understanding.
Claude Mythos 5.1 is the same underlying model as Fable 5.1, with looser cyber and biology safeguards. Fable 5.1 is generally available. Mythos 5.1 is restricted to US companies and individuals in Anthropic’s trusted access programmes.
The practical split on cyber: Fable 5.1 permits defensive vulnerability discovery, while exploit generation and some dual-use cyber work route to Opus models instead. A new Life Sciences Verification Program opens Mythos 5.1’s biology capabilities to scientific research, with wider enrolment for scientists said to be coming.
This is the same pattern as OpenAI gating GPT-6 Astra behind Critical-threshold cybersecurity controls. Two labs, the same quarter, both shipping frontier models with capability access split by verified identity rather than by payment.
Worth watching as a structural shift. The question of who gets the unrestricted model is becoming as consequential as which model is best.
One further note for enterprise buyers: Anthropic’s standard data retention policy does not permit zero data retention. Eligible enterprise customers can get ZDR until Enterprise Frontier Safeguards rolls out in phases from later in autumn 2026. If your compliance posture requires ZDR, confirm your status before building on this.
Claude Fable 5.1 vs GPT-6 Astra
The two flagships shipped two days apart and are priced identically to the dollar.
| Claude Fable 5.1 | GPT-6 Astra | |
|---|---|---|
| Released | 1 Sep 2026 | 3 Sep 2026 |
| Input / output per 1M | $10 / $50 | $10 / $50 |
| Cache reads | $0.25 | Higher |
| Context | 1M | 1.05M |
| Leads on | Agentic, independent indices | Computer use |
| Cost per Intelligence Index task | $2.37 to $7.63 | ~$1.67 |
Two labs landing on an identical rate card is not coincidence; both are pricing against each other rather than against cost. The differentiators are elsewhere.
Astra wins on cost per completed task because it spends fewer tokens reaching its score. Fable 5.1 wins on cache reads by 4x, which reverses the comparison for agent loops dominated by cache traffic. Astra leads computer use; Fable 5.1 leads Artificial Analysis’s indices and agentic benchmarks.
There is no universal answer, which is why our AI models compared roundup recommends routing rather than standardising on one.
Who Should Switch
| If you are | Do this | Why |
|---|---|---|
| Running agent loops on Fable 5 | Switch now | Cache reads cost a quarter as much |
| Using Claude Code heavily | Switch now | Cache-dominated workload, biggest saving |
| Running simple completions | Switch, low priority | Same price, modest gains |
| Using forced tool use | Read the migration guide first | Returns a 400 error on 5.1 |
| Running max effort by default | Drop to xHigh | Identical score, 28% cheaper |
| Cost-sensitive on routine tasks | Use low effort | $2.37 per task, still scores 47 |
| Needing computer use | Test GPT-6 Astra | Astra leads that specific category |
| Requiring zero data retention | Confirm eligibility first | Not standard policy until autumn 2026 |
Where It Sits in a Real Stack
Worth placing this against the wider field, because $10 and $50 per million is frontier pricing and most traffic does not need it.
Against open weights. GLM-5.3-Flash costs $0.15 and $0.50 per million, roughly a hundredth of Claude Fable 5.1 on output. For bulk classification, summarisation or extraction, paying frontier rates is hard to justify. The sensible architecture routes most traffic cheaply and escalates only hard tasks.
Against hosted inference providers. Groq and Cerebras serve open models at a fraction of this cost with far higher throughput, covered in our AI inference API comparison. They cannot run Claude Fable 5.1, because Anthropic does not license its weights to third parties.
Against running locally. No amount of hardware runs this model. Closed weights mean the only access is Anthropic’s API or a cloud partner, which is the structural trade-off our best local LLMs guide sets against open alternatives.
The realistic position for Claude Fable 5.1 is the top of a routing tier rather than the default for everything. Use it where long-horizon agentic work and hard reasoning justify the rate, and route the rest elsewhere.
Frequently Asked Questions About Claude Fable 5.1
Is Claude Fable 5.1 cheaper than Fable 5?
The list price is identical at $10 and $50 per million tokens. Cache reads dropped 75% to $0.25, which Anthropic estimates makes typical workloads about 25% cheaper and highly agentic work up to 45% cheaper.
Will switching break my code?
Possibly. Forced tool use now returns an error, and thinking block handling changed. Simple completions migrate cleanly; agent code needs the migration guide first.
What effort level should I use?
xHigh for quality work, since it matches Max at 53 on the Intelligence Index for $1.65 less per task. Low for routine or checkable work at $2.37. Max is rarely worth its 28% premium over xHigh.
What is Claude Mythos 5.1?
The same underlying model with looser cyber and biology safeguards, restricted to US organisations and individuals in Anthropic’s trusted access programmes. Fable 5.1 is the version generally available.
Can Claude Fable 5.1 do security work?
Defensive vulnerability discovery, yes. Exploit generation and some dual-use cyber work route to Opus models instead.
How does it compare to GPT-6 Astra?
Identical list pricing. Fable 5.1 leads agentic benchmarks, independent indices, and cache-read costs by 4x. Astra leads computer use and costs less per completed task because it uses fewer tokens. The right choice depends on workload shape.
What to Do in the First Week
A short sequence for anyone deciding whether to move to Claude Fable 5.1.
Check whether your bill is cache-dominated. Pull a week of usage and look at the ratio of cache reads to fresh input tokens. If cache reads are a large share, the 4x cut is worth acting on immediately. If they are negligible, this release changes little for you.
Grep your codebase for forced tool use. That is the breaking change most likely to be in production already, and it fails with a 400 rather than degrading gracefully.
Run your evaluation set at low and xHigh effort, not just max. Measure whether the six index points between them actually change your outputs. On many workloads they will not, and the cost difference is 2.5x.
Compare against Astra on your own prompts. Identical rate cards mean the published comparisons will not tell you which is cheaper for your work. Measure cost per completed task rather than per token.
Verdict
Claude Fable 5.1 is a more significant release than the unchanged price tag suggests, and the reason is a single footnote. Cutting cache reads by 75% takes a quarter to nearly half off agentic workloads without any change to how you use the model, which makes staying on Fable 5 for agent loops straightforwardly wasteful. Ranking first of 232 on independent aggregation, with agentic as its strongest category, backs up Anthropic’s positioning.
The honest caveats are the ones buried furthest down the announcement. Three breaking changes will return errors rather than degraded output on existing agent code, and forced tool use is the most likely to already be in your codebase. The effort tiers cost 3.2 times apart for six index points, with Max matching xHigh on quality while costing 28% more, which means the default many teams reach for is the worst value in the range. And Terminal-Bench-Science more than doubling in a point release is the kind of jump that deserves independent reproduction before anyone builds a decision on it.
Concrete next step: pull one week of API usage and look at what share of your spend is cache reads. That single number tells you whether this release saves you 45% or almost nothing, and it takes five minutes to find.
Sources and Further Reading
- Anthropic’s Claude Fable 5.1 announcement — the launch post, system card and migration guidance
- LLM Stats launch analysis — the self-reported benchmark table and cache multiplier breakdown
- Artificial Analysis effort-tier comparison, intelligence, speed and cost per task across all five effort settings
- Codersera migration guide, the three breaking changes and full pricing table
- BenchLM model profile, independent aggregate scoring, category rankings and safeguard routing detail
For related coverage on this site, see GPT-6 Astra for the rival that shipped two days later, AI models compared for where both sit against the field, LLM API cost optimization for reducing spend before switching models, and AI coding assistants for the tools this model powers.