Claude Fable 5.1 vs Opus 5: Benchmarks, Price, Verdict

Anthropic shipped Claude Fable 5.1 on 1 September 2026, six weeks after Claude Opus 5 landed on 24 July. The two sit at the top of the Claude lineup with identical context windows and identical output limits — and a price gap of exactly 2×.

That 2× is where most comparisons stop. It's also the least interesting number here, because Fable 5.1 inverts one line of the price sheet: it charges half what Opus 5 charges for prompt-cache reads. For a long-running agent re-reading a big cached prefix every turn, that single line changes the answer completely. Below is the arithmetic.

How do Claude Fable 5.1 and Opus 5 compare on specs?

Both models are pinned snapshots on the Claude API — claude-fable-5-1 and claude-opus-5. Everything below is from Anthropic's models overview and pricing page.

Two rows deserve a second look. The cache-read row is the only place Fable 5.1 undercuts Opus 5, and it does so by half. And Opus 5's fast mode bills at $10/$50 — exactly Fable 5.1's list price — for up to 2.5× the output speed. If you have budget for Fable 5.1 rates, a faster Opus 5 is a real alternative, not a consolation prize.

What do the benchmarks actually say?

Anthropic published a same-harness comparison across three models in its Fable 5.1 announcement. These are vendor-reported numbers — Anthropic ran them — but they at least put all three models on the same evals.

That table is the whole argument. Fable 5.1 wins all five — but outside one benchmark, by 1.4 to 3.5 points. The single blowout is Terminal-Bench-Science, where it nearly doubles Opus 5 (52.6% vs 29.0%) — a scientific-research agentic eval, not a general coding one. That is what Anthropic means in positioning Fable 5.1 as a research and long-horizon model rather than a general upgrade.

Independent measurement agrees on the size of the gap. Artificial Analysis scores Fable 5.1 at 66 on its Intelligence Index against 63 for Opus 5 — three points, consistent with the non-science benchmarks above. Artificial Analysis also measures Fable 5.1 as faster in raw throughput (66 tok/s vs 56 tok/s) but dramatically slower to start answering: time-to-first-token of 285 seconds versus 82 seconds for Opus 5. Always-on adaptive thinking has to happen somewhere, and it happens before you see anything.

On human preference, LMArena as of 2 September 2026 ranked Fable 5.1-max third at 1504±11 (2,906 votes) and Opus 5-high ninth at 1493±5 (35,174 votes). Treat that with caution — 2,906 votes and a ±11 interval means Fable 5.1 has not cleanly separated from Opus 5 yet. Note also that Opus 5 at high effort outscores Opus 5 at max: on open-ended chat, more thinking is not automatically better.

One honest gap: neither model's SWE-bench Verified score appeared in the launch materials or documentation we could verify. Anthropic used Terminal-Bench 4.0 and CursorBench 3.2.0 as its coding-agent evals instead, so those are what we've quoted.

Why does Fable 5.1 cost 2× on list but half on cache reads?

Anthropic's pricing page carries a footnote that does more work than the whole rest of the table: "Cache hits and refreshes on Claude Fable 5.1 and Claude Mythos 5.1 are priced at 0.025x the base input price. All other models use the standard 0.1x multiplier."

The mechanism matters more than the two dollar figures. Fable 5.1 sits on a 2.5% cache multiplier; Opus 5 — like every other Claude model, including Fable 5 — sits on the standard 10%. Because Fable 5.1's base input is $10 and Opus 5's is $5, that four-fold multiplier advantage more than cancels the two-fold base-price disadvantage, and the arithmetic lands on $0.25 against $0.50. Fable 5.1 is half price on the one token category that dominates agentic workloads. It's also a 4× cut from Fable 5's own $1.00/MTok.

What does a stateless request cost on each?

A one-shot call with no caching — 10,000 input tokens, 2,000 output tokens:

• Fable 5.1: (10,000 × $10 + 2,000 × $50) ÷ 1M = $0.20
• Opus 5: (10,000 × $5 + 2,000 × $25) ÷ 1M = $0.10

Clean 2×. If your workload looks like this — stateless classification, one-off generation, RAG with a fresh context each time — the price gap is exactly what the sticker says and there is no clever argument to be had.

What does a cached coding agent actually cost?

Now a realistic long-running agent: a 200,000-token cached prefix (system prompt, tool definitions, repository context), 50 turns, each turn adding 2,000 new input tokens and producing 1,500 output tokens, using the 5-minute cache.

The gap has collapsed from 2× to 1.13×. Fable 5.1 costs 13% more, not 100% more. The cache-read line alone hands back $2.50, which nearly cancels the $1.88 output premium and the $1.25 cache-write premium combined. For a 3-point intelligence gain, a 13% premium is a genuinely different decision from a 100% one.

When does Fable 5.1 become genuinely cheaper?

Push the same logic further. Take a workload dominated by cache reads: a 1M-token cached document corpus held in the 1-hour cache, queried 500 times, each query adding 100 input tokens and producing 200 output tokens.

Fable 5.1 is 43% cheaper — a $112 saving on a $263 job, while also being the stronger model. The inversion is real, not a rounding artefact.

The general break-even, in tokens: Fable 5.1 wins on cost when cache_reads > (20 × new_input) + (25 × cache_writes) + (100 × output). The practical version of that rule is simpler — your cache-read volume has to exceed roughly 100× your output volume before the cheaper reads outweigh the doubled output price. Chatty agents that write a lot of code never get there. Retrieval and analysis loops over a large fixed corpus get there easily.

One caveat on Anthropic's own framing. The announcement says costs drop "approximately 25% for typical workloads and up to 45% for highly agentic tasks" — that is Fable 5.1 versus Fable 5, not versus Opus 5. It is easy to misread as a claim about this matchup. It isn't one.

What about batch and fast mode?

The Batch API takes 50% off input and output on both models, and stacks with cache multipliers. That produces a result worth knowing:

• Fable 5.1 on batch is $5 / $25 — identical to Opus 5 at list price. If your workload tolerates asynchronous processing, you can run the stronger model for what the cheaper model costs synchronously.
• Opus 5 on batch drops to $2.50 / $12.50, so the 2× ratio holds if you batch both.
• Fast mode is Opus 5 only. At $10 / $50 it costs the same as Fable 5.1 list for up to 2.5× output speed — the right trade when latency, not intelligence, is your constraint.

Which model is better for coding and agentic work?

Fable 5.1, but by less than the price implies. It leads Terminal-Bench 4.0 55.8% to 52.3% and CursorBench 3.2.0 73.4% to 70.0%. Anthropic's own Opus 5 announcement claimed Opus 5 lands "within 0.5% of Fable 5's peak score" on CursorBench 3.2 at maximum effort "but at half the cost per task" — and Fable 5.1 has only moved that needle 3.4 points.

Where Fable 5.1 pulls decisively ahead is duration. Anthropic lists its gains as concentrated in agentic coding over sessions that run for hours, multistep research that follows up on what it finds, long-context reasoning across the full 1M window, and computer use with recovery from failed steps. If your agent runs for twenty minutes, Opus 5 is fine. If it runs for four hours across a large refactor, the compounding error rate is what you're paying to reduce.

Two countervailing details sit in Anthropic's migration notes. Fable 5.1's parallel tool calling is more variable — it may issue one tool call per turn where Fable 5 batched several, costing round trips and wall-clock time. It is also "more likely to rewrite the entire file than make a targeted edit," burning output tokens at $50/MTok. Both are prompt-fixable, but neither shows up in a benchmark score and both work against the cost model above.

Where is each model genuinely worse?

Fable 5.1's real drawbacks:

• Forced tool use returns a 400 error. tool_choice of any or tool is rejected outright: tool_choice: type "tool" and "any" are not supported for this model. The cause is structural rather than an oversight — thinking is always on, and a forced call would skip it, pushing the model's working-out into the tool arguments and degrading their quality. auto and none still work, and Anthropic's suggested replacements are strict tool use or structured outputs. If your integration depends on pinned tool selection, migrating is code work, not a string swap.
• Thinking cannot be disabled. thinking: {"type": "disabled"} errors. You pay for reasoning tokens at $50/MTok on every call, including trivial ones.
• Editing earlier turns invalidates thinking blocks. Thinking blocks are bound to both the model that produced them and the conversation prefix, so changing the system prompt, the tools array, or any earlier message errors the next request. Harnesses that rebuild the messages array must become append-only. The check is enforced for accounts created on or after 31 August 2026.
• Latency. Anthropic labels it "Slower"; independent measurement puts time-to-first-token at 285 seconds against 82 for Opus 5.
• Data retention. Fable 5.1 carries 30-day retention and is not available under zero data retention unless Anthropic expressly authorises it. For some regulated buyers that ends the conversation before price does.
• No fast mode. At low effort it answers from memory more often instead of calling search tools.

Opus 5's real drawbacks:

• A lower ceiling on the hardest reasoning. Terminal-Bench-Science 29.0% vs 52.6% is not a rounding difference — on novel scientific work, Opus 5 is not in the same class.
• Cache reads cost double. $0.50 vs $0.25/MTok is the one line where the cheaper model is the expensive one, and it dominates long agent loops.
• It trails on every published head-to-head benchmark, and its knowledge cutoff is a month earlier (May vs June 2026).

Which should you actually pick?

Anthropic's own guidance is unusually blunt for a vendor selling the more expensive model. Its Choosing a model page states plainly: "Most workloads start with Claude Opus 5." The escalation rule is to move to Fable 5.1 only when your evals on Opus 5 at higher effort still fall short.

That is the decision rule, and the benchmark spread supports it. Tuning the effort parameter is usually a better lever than switching models, and it's free.

One architectural gate sits upstream of any cost calculation. If your agent pins tool selection with tool_choice: {"type": "any"} or a named tool, Fable 5.1 is not a drop-in. Router layers that force a classification call, extraction pipelines mandating a schema tool, and state machines requiring a call at a known step all break with a 400 — as does any harness that mutates conversation history between turns. No benchmark or price table surfaces that. Check it before you model the tokens.

FAQ

Is Fable 5.1 better than Opus 5?

Yes, on every published head-to-head benchmark — but usually narrowly. Fable 5.1 leads Terminal-Bench 4.0 55.8% to 52.3%, CursorBench 3.2.0 73.4% to 70.0%, and Artificial Analysis scores it 66 to 63 on intelligence. The one decisive gap is Terminal-Bench-Science, at 52.6% versus 29.0%. Outside scientific and long-horizon research work, expect three points, not a generational leap.

Is Fable 5.1 worth 2× the price of Opus 5?

For most workloads, no — Anthropic itself says to start with Opus 5. But 2× only applies to uncached calls. In a realistic cached agent loop the premium falls to about 13%, and in cache-read-dominated workloads Fable 5.1 is outright cheaper. Work out your own token mix before accepting the sticker ratio as your actual cost.

Which is better for coding?

Fable 5.1, by 3.5 points on Terminal-Bench 4.0 and 3.4 on CursorBench 3.2.0. For most coding, that gap is not worth double the list price — raise Opus 5's effort parameter first. Fable 5.1 earns its keep on multi-hour autonomous sessions, large refactors, and migrations where compounding errors cost more than tokens.

Which is cheaper with prompt caching?

Fable 5.1 charges $0.25/MTok for cache reads against Opus 5's $0.50 — 2.5% of base input versus the standard 10%. In a 200k-prefix, 50-turn agent loop that cuts Fable 5.1's premium from 100% to about 13%. In a cache-read-dominated workload — roughly 100× more cache reads than output tokens — Fable 5.1 becomes 40%+ cheaper outright.

What's the context window on each?

Both offer a 1M-token context window with 128K max output tokens, and both bill the full window at standard per-token rates — a 900k-token request costs the same per token as a 9k one. Both use the tokenizer introduced with Opus 4.7, so 1M tokens is roughly 555,000 words. There is no context-window advantage either way.

Should I switch from Opus 5 to Fable 5.1?

Only if your evals on Opus 5 at higher effort still fall short — that's Anthropic's own escalation rule. Budget migration work if you do: forced tool use now returns a 400 error, thinking cannot be disabled, and editing earlier turns invalidates thinking blocks, so any harness that rebuilds its message history needs changing first.

The bottom line

Fable 5.1 is the better model and Opus 5 is the better default. A three-point intelligence gap does not justify a 2× bill for stateless work, and Anthropic's own documentation says as much. What genuinely changes the calculation is the cache-read inversion: at $0.25 against $0.50/MTok, the expensive model gets cheaper the more you cache, and past roughly 100× cache reads to output tokens it wins on both quality and cost.

So measure your token mix, not the headline price. Pull cache-read, cache-write, new-input and output totals from a week of production traffic, run them through the four line items above, and the answer falls out. Most teams will find Opus 5 at higher effort is the efficient choice — the ones running hours-long agents over a large cached corpus will find the opposite, decisively.

Related reading: our complete guide to Claude Fable 5.1, the earlier Opus 5 vs Fable 5 comparison, the Claude Opus 5 launch guide, and how Fable 5 usage credits work.