Claude Fable 5.1 vs Opus 5, Fable 5 and GPT-5.6 Sol

Claude Fable 5.1 and Mythos 5.1 both shipped on September 1, 2026. Back in July I wrote a comparison built around Opus 5 and ended it with “worth revisiting in a few weeks.” This is that revisit.

For day-to-day work, the pricing structure and the breaking changes matter more than the benchmark deltas. And there are several places where the vendor table and third-party measurements point in opposite directions.

Claude Fable 5.1

claude-fable-5-1

Input

$10 / MTok

Output

$50 / MTok

Cache read

$0.25 / MTok

Context

1M

Claude Opus 5

claude-opus-5

Input

$5 / MTok

Output

$25 / MTok

Cache read

$0.50 / MTok

Context

1M

Claude Fable 5

claude-fable-5

Input

$10 / MTok

Output

$50 / MTok

Cache read

$1.00 / MTok

Context

1M

GPT-5.6 Sol

gpt-5.6-sol

Input

$5 / MTok

Output

$30 / MTok

Cache read

—

Context

1.05M

Specs

Base input and output prices are identical to Fable 5: $10 / $50. Context is 1M both by default and at maximum. The one thing that changed is the cache read price. The knowledge cutoff of June 2026 is the most recent of the four.

Batch API is half price at $5 / $25. Model IDs: anthropic.claude-fable-5-1 on Bedrock, claude-fable-5-1 on Vertex and Foundry.

Cache reads cut to a quarter

This is the most substantive change in the release.

Cache reads go from $1.00 to $0.25 / MTok — a 75% cut. The docs put it as: cache reads on these models cost 0.025x the base input price, versus 0.1x on other Claude models.

Cache writes are unchanged: $12.50 for 5 minutes, $20 for 1 hour. The 512-token cache minimum is unchanged too. So the saving only lands if you re-read a long prefix many times.

Anthropic quotes two figures:

• Typical workloads: about 25% cheaper than Fable 5
• Highly agentic workloads: up to 45%

The announcement page names no benchmark or dataset behind either number. One third-party analysis attributes them to four weeks of blended August 2026 billing, but I could not verify that claim against any Anthropic page. That analysis breaks it down as:

One-shot calls save nothing, which follows directly from never hitting the cache.

The measurement that points the other way

Artificial Analysis measured cost per task at max effort as $3.69 for Fable 5.1 against $3.14 for Fable 5 — roughly 20% more expensive.

The reason is output volume. Running the same index, Fable 5.1 emitted 140M output tokens against Fable 5’s 83M. That is 1.7x, and output bills at $50/MTok with no discount applied.

Both claims are true about different halves of the bill. Anthropic is describing input-side cache economics; AA is describing output volume. Which one applies depends on whether your workload re-reads a lot of input or generates a lot of output. Per the same third-party analysis, Fable 5.1 undercuts Opus 5 somewhere around 140,000–215,000 cached tokens per turn, with the crossover point depending on whether you count cache writes.

Three breaking changes

Moving Fable 5 code straight to 5.1 will break in these places.

1. Forced tool use is gone

tool_choice with {"type":"any"} or {"type":"tool","name":...} now returns 400.

1
2

400 invalid_request_error
tool_choice: type "tool" and "any" are not supported for this model.

The same restriction is enforced on count_tokens. auto and none still work. Pinning a specific tool to force structured output is a common pattern, and all of it hits this.

2. Thinking blocks are bound to a model

Fable 5.1 can read thinking blocks produced by earlier models. The reverse does not hold — no earlier model can read Fable 5.1’s thinking blocks.

Unreadable blocks are dropped silently and not billed, which also means you get no signal that anything was dropped. Sending the thinking-binding-controls-2026-08-01 header surfaces them in input_transformations. Worth enabling if your pipeline mixes models.

3. Editing an earlier turn invalidates thinking blocks

Anthropic calls this “preserved thinking.” Any of the following returns 400:

• Rebuilding system or tools
• Reordering or removing an earlier turn
• Injecting per-turn text and then deleting it

1

400 The block is bound to a different conversation

This is enforced for accounts created on or after August 31, 2026. Older accounts only hit it if they explicitly set thinking.block_binding.prefix_mismatch_behavior. Implementations that reassemble conversation history on every turn run straight into this.

Note that Mythos 5.1 does not run this check, even though it is the same generation.

What was added

• Per-message effort mid-conversation — beta mid-conversation-output-config-2026-07-01. Does not invalidate the prompt cache. Available on Fable 5.1, Mythos 5.1 and Opus 5
• Turn-scoped system messages — clear_at: "next_user_message", beta mid-conversation-system-clear-at-2026-08-21
• thinking.display: "updates" — readable progress notes between tool calls, beta thinking-display-updates-2026-08-18
• Content provenance — a statistical text watermark on all output, plus C2PA Content Credentials on images and video retrieved through the Files API

Thinking is adaptive and always on. Both {"type":"disabled"} and budget_tokens return 400. Opus 5 at least allowed disabling thinking at effort high or below; here it cannot be turned off at all. The API default effort is high.

Permitted refusal-fallback targets are Opus 4.8 and Opus 5. When a fallback fires, the prompt-cache cost of switching is refunded as credit.

Before reading the benchmarks

The same caveats from the July post apply, plus one more that is specific to this release.

Anthropic’s launch table was run with production safeguards enabled, and tasks the safeguards intervened on were completed by a different model. That is stated in a footnote:

On tasks where these safeguards intervened, Fable 5.1 and Fable 5 scored a zero on OSWorld 2.0, and Fable 5 scored a zero on AutomationBench. In all other interventions from our safeguards, cybersecurity tasks were completed by Claude Opus 4.8, and biology tasks were completed by Claude Opus 5.

So the headline numbers include model substitution. That is a scaffold difference, not a pure single-model measurement. To Anthropic’s credit this is disclosed in their own footnote rather than hidden.

A few more conditions:

• Terminal-Bench-Science 0.1 carries a standard error of ±3.5–4.5 points per model. The public leaderboard reports Opus 5 at 30.0% and Fable 5 at 21.4%; Anthropic’s own run reproduces them at 29.0% and 24.7%. Anthropic describes both as within noise
• OSWorld 2.0 uses the benchmark authors’ August 2026 task release. Because the task files differ from earlier releases, these numbers are not directly comparable to previously published OSWorld 2.0 results — which is why no competitor score appears in that row
• The table has no Sonnet 5 column and no Gemini column

Anthropic’s launch table

All vendor-reported.

On Terminal-Bench 4.0, Mythos 5.1 scores 60.9%, above Fable 5.1’s 55.8%. It is the only row where Anthropic broke Mythos out separately.

Terminal-Bench-Science stands out: 24.7% to 52.6% is more than double. AutomationBench nearly doubles as well, 17.1% to 31.4%. In July I wrote that Opus 5 had pulled ahead on AutomationBench at 26.0%; two months later that lead is gone.

The other rows move much less. HLE is up 3.1 points, CursorBench 2.9. The gains concentrate in the agentic and terminal categories; everything else is incremental.

No SWE-bench this time

Anthropic published neither SWE-bench Verified nor SWE-bench Pro for Fable 5.1.

Search results will hand you 95.0% and 80.3% as if they were Fable 5.1 numbers. They are Fable 5 numbers. So are ARC-AGI-1 98.5%, ARC-AGI-2 89.2%, and MMMU 74.3%. Aggregator sites have the versions mixed up. One site lists Fable 5.1 SWE-bench Pro at 81.2% with no source and no methodology, so I left it out.

There is likewise no official Fable 5.1 figure for GPQA Diamond, AIME, BrowseComp or MRCR.

Third party — Artificial Analysis

On Intelligence Index v4.1.1, Fable 5.1 scores 66, ranked 1st of 192 models.

One caveat on reading that table. Fable 5 scored 64.9 and ranked 1st on this index at its own launch. Its current page shows 62. AA appears to re-baseline across index versions, so 66 against 64.9 is not a like-for-like 1.1-point improvement.

The measurement configuration matters too. AA ran it as “Adaptive Reasoning, Max Effort, Default Fallback” — max effort with fallback enabled.

AA’s own commentary on the same page is not flattering: “particularly expensive when comparing to other models of similar price,” and “slower than average and very verbose.”

285 seconds to first token means nearly five minutes of nothing.

Third party — Vals AI

Vals Index 67.87%, ranked 1st of 51, run on September 1.

Individual benchmarks are mostly first place: ProofBench 100%, MMLU Pro 92.38%, MMMU Pro 90.64%, LiveCodeBench 90.52%, MedScribe 91.29%. Terminal-Bench 2.1 is second at 85.02%.

Scoring fallbacks as failures flips the ranking

From Vals’ own methodology note:

Refusals remain frequent on bio- and cyber-adjacent tasks, so we ran Fable 5.1 with Claude Opus 5 and Claude Opus 4.8 as server-side fallbacks.

They also published the numbers with fallback-assisted tasks counted as failures.

The adjusted 66.85% falls below Opus 5’s 67.21%. First place hinges on one accounting decision. This is the same issue as the model substitution in Anthropic’s footnote.

One conspicuously weak result

Harvey’s Legal Agent: 6.67%, ranked 18th of 55. Nearly everywhere else Fable 5.1 lands in the top three. Fable 5 ranked 6th of 55 on the same benchmark, so this is a regression from the previous version.

Cost and time are steep as well: $28.40 per test at 72 min 25 s latency. Fable 5 was $28.80 at 37 min 51 s — about the same price, but 1.9x the time.

Third party — CodeRabbit code review

Measured by an actual code review product.

Recall is 0.9 points below Fable 5, precision 4.5 points above. That reads as the same ability to find known issues with fewer false positives. Total comments dropped to 166, 87 fewer than Fable 5, and nitpick comments fell 70.2% to 79.

The cost is time: 18 min 38 s per task, 6 min 6 s slower (+48.7%) than Fable 5.

One counter-intuitive result: low effort beat high effort. 61.0% recall in 18:38 at low against 57.1% in 21:36 at high. Slower and worse.

The July Opus 5 post found the same shape, where max scored below xhigh. Turning effort up is not automatically the right move. One analysis notes that effort level names do not map to the same amount of thinking across models, so effort values tuned on an earlier model do not carry over. The same source reports that at medium, Fable 5.1 roughly matches Fable 5 at lower cost.

Regressions Anthropic documented itself

The “Changed from Claude Fable 5” section is unusually candid — it lists what got worse alongside what got better.

• Parallel tool calling is more variable. It may issue one tool call per turn where Fable 5 batched several. Costs tokens, round trips and wall clock. Answer quality is not affected
• Fewer progress updates during long tool runs, especially at higher effort
• Answers from memory more often at low effort — it calls search and retrieval tools less
• Denser prose — longer sentences, fewer paragraph breaks
• Less formatting in chat — bold, headers and lists used less than earlier Claude models
• Unmarked quotations in summaries — it reproduces source passages without marking them as quotes
• Whole-file rewrites for small edits — more output tokens and more time

That last one lines up with the 1.7x output token count above. The unmarked-quotation behaviour is the one worth watching.

Mythos 5.1

Shipped the same day, but positioned differently.

• The docs describe it as having “the same capabilities as Claude Fable 5.1” — same specs, same pricing, including the $0.25/MTok cache read
• Invite only, restricted to Project Glasswing participants. Access runs through the Cyber Verification Program and the Life Sciences Verification Program
• One outlet reports access is limited to U.S. companies and individuals, with open enrollment for Life Sciences “coming soon”
• It does not run the preserved-thinking check that Fable 5.1 runs
• Cyber safeguards are relaxed for vetted defenders

Per summaries of the system card, it has the strongest cyber capability of any Anthropic model and substantially outperforms Opus 5 on nearly all cyber evaluations, including ExploitBench, OSS-Fuzz, Firefox 147 and ExploitGym. Chem-bio capability sits at CB-1, below the CB-2 threshold. I could not obtain the numeric scores — the system card PDF exceeded the fetch size limit.

The science claims from the announcement: protein binders with binding affinities 10x higher than the best competition designs, a ~50% hit rate across 12 targets against a typical 10–15%, deep-learning model optimization with speedups up to 2.5x and an estimated 30–60% GPU cost reduction, and a Venus elevation map at 2–3 km resolution against a prior 10–20 km. A footnote concedes that one competing entry was comparable to their best binder.

Community reaction

The dominant complaint in the Hacker News thread is about prose style, not capability.

Reports include purple prose, invented jargon, sycophantic phrasing, and fabricated document references with made-up dates in code comments. Several commenters say they reverted to Opus 4.8 or switched to GPT-5.6 Sol over readability.

There are also complaints about safety guardrails tripping on routine CRUD web development, and about the absence of a default ZDR mode making it hard to adopt in academic settings. That one checks out: Fable 5.1 requires 30-day retention and is not available under ZDR unless expressly authorized.

A GitHub issue alleging a broader quality regression across the generation also exists, but I did not read the issue body, so I am not summarizing its claims here.

Verdict

Fable 5.1 is not a broad capability jump. It is a gain on specific axes plus a pricing change.

The terminal and agentic categories genuinely moved: Terminal-Bench-Science 24.7% to 52.6% and AutomationBench 17.1% to 31.4% are both roughly double. Meanwhile HLE gained 3.1 points, CursorBench 2.9, and code review recall actually fell 0.9. The variance by area is large.

The first-place claims are conditional. Fable 5.1 leads both the AA Intelligence Index (66) and the Vals Index (67.87%), and both are accurate as run. But Vals drops to 66.85% — below Opus 5 — when fallbacks count as failures, and Anthropic’s own table had Opus 4.8 and Opus 5 completing refused tasks. “Fable 5.1 plus Opus fallback is first” is the more precise statement than “Fable 5.1 is first.”

Cost moves in two directions. The 75% cache-read cut is real, but it applies to the input side. If the model emits 1.7x the output tokens, the total goes up instead — AA measured 20% more per task. Long reused prefixes win; long outputs lose.

Speed clearly regressed. Vals 72 minutes (1.9x), CodeRabbit +48.7%, AA 285 seconds to first token. No measurement showed it getting faster. That is a lot to absorb for interactive use.

What to use when.

Check before migrating.

1. Anywhere using tool_choice with any or tool — returns 400. This is the most common thing that will break
2. Code that reassembles conversation history each turn — 400 from preserved thinking. Enforced for accounts created on or after August 31, 2026
3. Pipelines that mix models — enable thinking-binding-controls-2026-08-01 so dropped blocks become visible
4. Re-tune effort. Values from earlier models do not carry over. Start from medium; for code review there is a measurement showing low did better
5. Code using thinking: disabled — it cannot be turned off at all now

Most of these numbers are vendor-reported or were measured on launch day. Same situation as the Opus 5 post in July. Expect the picture to shift as independent measurements accumulate.

References

• Anthropic announcement — https://www.anthropic.com/claude-fable-and-mythos-5-1
• Fable 5.1 docs — https://platform.claude.com/docs/en/models/fable-5-1/overview
• What’s new in Fable 5.1 — https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1
• Mythos 5.1 docs — https://platform.claude.com/docs/en/models/mythos-5-1/overview
• Artificial Analysis — https://artificialanalysis.ai/models/claude-fable-5-1
• Vals AI — https://www.vals.ai/models/anthropic_claude-fable-5-1
• Vals Index — https://www.vals.ai/benchmarks/vals_index
• CodeRabbit review — https://www.coderabbit.ai/blog/fable-5-1-model-review
• VentureBeat — https://venturebeat.com/technology/anthropics-claude-fable-5-1-and-mythos-5-1-arrive-with-a-75-cost-reduction-for-fable-cache-reads