How much more does Claude Fable 5.1 cost than Opus 5.5?

On list price, 2.5x for input, output and cache writes, but only 1.25x for cache reads. On our real Claude Code token mix that came to 2.13x to 2.31x. Per turn, as actually run, Fable cost $0.284 against $0.064, about 4.4x, because it writes more output at its default high effort.

Is Fable 5.1 better than Opus 5.5 for coding?

Not on agentic coding benchmarks. Anthropic reports Terminal-Bench 4.0 at 66.4% for Opus 5.5 and 55.8% for Fable 5.1, and Snorkel's independent run has 64.8% vs 57.9% with Fable's run costing 1.3x more. Fable leads on general reasoning indexes.

Does Fable use up Claude Max limits faster?

Yes. Anthropic says Fable models use limits faster and can take up to 50% of a Max weekly limit, but does not publish the rate. At list-price ratios, one Fable turn costs about four Opus 5.5 turns of quota.

When should I use Fable 5.1 instead of Opus 5.5?

For few-turn, high-stakes reasoning such as architecture decisions or hard root-cause analysis. Keep Opus 5.5 as the default for coding agents, and keep Fable to about 10% of a Max week.

Claude Fable 5.1 vs Opus 5.5: the real price multiple from our Claude Code logs

Updated October 10, 2026 · first published October 10, 2026

Quick answer: Claude Fable 5.1 lists at $10 input and $50 output per million tokens. Opus 5.5 lists at $4 and $20. That is the 2.5x everyone quotes. It is right on the headline and wrong on the bill: the cache...

Claude Fable 5.1 lists at $10 input and $50 output per million tokens. Opus 5.5 lists at $4 and $20. That is the 2.5x everyone quotes. It is right on the headline and wrong on the bill: the cache read, where an agent spends most of its tokens, is only 1.25x. We priced every Fable and Opus 5.5 turn in four full weeks of our own Claude Code logs to see what the premium really is.

Price on our real token mix: 2.13x for Opus-style turns, 2.31x for Fable's own turns. Not 2.5x.
Cost per turn, as actually run: a Fable 5.1 turn cost $0.284, an Opus 5.5 turn $0.064. That is 4.4x, because Fable writes about 2.7x more output per turn on its default high effort.
Coding benchmarks favour Opus 5.5: Anthropic's own Terminal-Bench 4.0 figures put Opus 5.5 at 66.4% and Fable 5.1 at 55.8%. Snorkel's independent run: 64.8% vs 57.9%, with Fable's run costing 1.3x more.
Reasoning benchmarks favour Fable: Artificial Analysis scores Fable 5.1 at 66 on its intelligence index at max effort, the top Claude score.
On a Max plan: Fable can take up to half your weekly limit and burns it faster. We kept it to 0% to 10% of each week's usage.

Fable 5.1 price as a multiple of Opus 5.5, by token type and on our real turns

What Anthropic claims

These are vendor claims, quoted as published or as reported by launch coverage. Our measurements come after.

Claim	Source
Fable 5.1 lists at $10 input, $50 output, $0.25 cache read, $12.50 5-minute write and $20 1-hour write per million tokens.	API pricing page
Opus 5.5 lists at $4 input, $20 output, $0.20 cache read, $5 5-minute write and $8 1-hour write.	API pricing page
Fable 5.1 is about 25% cheaper than Fable 5 for typical workloads and up to about 45% cheaper for agentic ones, mostly because the cache read dropped from $1.00 to $0.25.	Launch coverage (Constellation Research, NYU Shanghai RITS)
Opus 5.5 performs close to Fable 5.1, and at default effort beats Opus 5 at max effort for about a fifth of the cost.	Opus 5.5 launch post
Terminal-Bench 4.0: Opus 5.5 66.4% at xhigh effort, Fable 5.1 55.8%.	Opus 5.5 launch post
Fable 5.1 scores 60.9% on Humanity's Last Exam without tools and 65.0% with tools. High effort is its default.	Launch coverage
On Max plans, Fable models can use up to 50% of your weekly limit, have their own weekly limit, and use limits faster than other models. On Pro, Fable runs on pay-as-you-go usage credits.	Claude Fable models on your plan
Claude Code defaults to Opus 5.5 on paid plans. Fable is never the default; you pick it with /model.	Claude Code model configuration docs

What Anthropic does not publish: how much faster Fable drains a plan limit. The plan pages say "faster" and stop there.

Why the real multiple is 2.1x to 2.3x, not 2.5x

An agent turn in Claude Code re-reads the whole conversation from cache, then writes a little new output. Across our 22,411 Opus 5.5 turns, the average turn read about 94,000 cached tokens and wrote about 551 output tokens. Cache reads are 100x cheaper per token than output, but there are 170 times more of them, so they carry real weight in the bill. Five token types, five prices:

Token type, per million	Opus 5.5	Fable 5.1	Multiple
Input	$4.00	$10.00	2.5x
Output	$20.00	$50.00	2.5x
Cache write, 5 minutes	$5.00	$12.50	2.5x
Cache write, 1 hour	$8.00	$20.00	2.5x
Cache read	$0.20	$0.25	1.25x

Re-price the exact tokens of our Opus 5.5 turns at Fable rates and the bill goes from $1,436 to $3,061: 2.13x. Do the reverse for our Fable turns and you get 2.31x; Fable turns lean more on output and 1-hour cache writes, so less of their bill gets the cheap cache-read discount. The more cache-heavy your agent, the further below 2.5x you land. A pure chat workload with no cache sits at the full 2.5x.

Per turn it is 4.4x, and price is only half of that

Average cost of one Claude Code turn, our logs, Sep 11 to Oct 9, 2026

Per turn, our logs	Opus 5.5	Fable 5.1
Turns	22,411	486
Output tokens per turn	551	1,461
Cache read tokens per turn	about 94,000	about 94,000
1-hour cache write tokens per turn	3,939	9,383
Cost per turn at own list price	$0.064	$0.284
Total, four weeks	$1,436	$138

Context size was identical, about 94,000 cached tokens per turn on both. What differed was behaviour. Fable wrote 2.7x more output per turn and 2.4x more fresh cache. That is high effort, Fable's default, doing more thinking per step. Price explains about 2.3x, behaviour the rest, for 4.4x per turn.

One honest caveat: we picked Fable for the hard turns. Some of that extra output is the task, not the model. Even so, the per-turn gap is the number your bill sees.

Raw data: fable-5-1-vs-opus-5-5-turns.csv.

What the extra money buys

Benchmark	Opus 5.5	Fable 5.1	Who ran it
Terminal-Bench 4.0	66.4% (xhigh)	55.8%	Anthropic
Terminal-Bench 4.0, score	64.8%	57.9%	Snorkel, independent
Terminal-Bench 4.0, cost of the run	$4.7k, 8.0B tokens	$6.2k, 2.7B tokens	Snorkel, independent
Intelligence index, cost per task	not on the same page	66 at max effort, $3.76; 65 at xhigh, $2.72	Artificial Analysis

On agentic coding, Opus 5.5 wins on score and on cost. Snorkel's run shows the pattern from our logs in reverse: Fable used a third of the tokens but still cost 1.3x more, because each of its tokens is priced higher. On Artificial Analysis's general reasoning index, Fable 5.1 at max effort is the top Claude score, ahead of Opus 5 at max effort (63, $2.34 per task). Artificial Analysis lists Opus 5.5 separately at 51 to 56 across medium to xhigh effort, for $1.34 to $3.46 a task. We have not seen those on the same comparison page, so treat that gap as indicative.

What it does to a Claude Max week

We use the full Claude Max 5x weekly cap every week. Priced at today's list prices (Opus 5 and Sonnet 5 turns repriced to Opus 5.5 and Sonnet 5.5), the four weeks were worth $419, $761, $873 and $658. Fable's share of that was $39, $34, $0 and $65: 9.3%, 4.5%, 0% and 9.9% of each week. Anthropic lets Fable take up to 50% of a weekly limit and says it uses limits faster, but not by how much. If the plan meters near list price, every Fable turn costs about four Opus 5.5 turns of quota. Spend half the week on Fable and you lose about 40% of the turns you would otherwise get.

When Fable 5.1 is worth 2.5x

Default to Opus 5.5 for coding agents. It wins Terminal-Bench by 7 to 11 points and costs about a quarter as much per turn. Claude Code already defaults to it.
Switch to Fable for one-shot hard reasoning: an architecture call, a gnarly root-cause analysis, a spec review. Few turns, high stakes, and it leads on reasoning benchmarks.
On the API, the cache narrows the gap. A long-context agent with a high cache hit rate pays about 2.1x, not 2.5x. Lower effort narrows it further.
On a Max plan, cap it. Keep Fable to about 10% of the week, as we did. Past that it starts eating turns you need for Opus.

For the full plan comparison against Codex, see Claude vs Codex in October 2026. For the cache-hit maths behind Fable pricing, see the Fable 5.1 cost model.

What this does not show

This is one heavy user's logs. Fable ran 486 turns against 22,411 for Opus 5.5, and we chose Fable for harder turns, so its per-turn cost is biased upward. Dollar figures are API-list value, not money we paid on a flat plan. Some Fable figures come from launch coverage quoting Anthropic, because we could not find Anthropic's own Fable 5.1 launch post. Prices agree with Anthropic's pricing page.

Sources

Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →

Back to finopsllm.com