Claude Opus 5.5 Guide: Cheaper Than Opus 5, Close to Fable 5.1

Successor models usually cost the same or more. Claude Opus 5.5 went the other way. Anthropic released it on 22 September 2026 at $4 per million input tokens and $20 per million output, down from $5 / $25 for Opus 5, and published a table in which it beats both Opus 5 and the far more expensive Claude Fable 5.1 on every row.

Cheaper and better is a strong claim, so it deserves a careful read. This guide covers what the numbers actually say, the default that changed underneath you, the four breaking changes, and how to check the cost claim on your own work.

Cheaper than the model it replaces

Opus 5.5 is the first model in the Claude 5.5 family. It keeps Opus 5's 1M-token context, 128K maximum output and tokenizer, so token counts carry over. The price changes, per Anthropic's pricing page, are broader than the headline:

• Input and output — $4 and $20 per million tokens, 20% lower than Opus 5.
• Cache reads — $0.20 per million, 60% lower than Opus 5's $0.50, and a deeper discount relative to base input than most Claude models get.
• Cache writes — $5 per million for the five-minute cache, against $6.25.
• Batch — $2 / $10 per million tokens on the asynchronous Batch API.
• Fast mode — $8 / $40, up to about 2.5 times normal speed.

Anthropic also says the model costs about 40% less than Opus 5 on typical workloads at default settings and generates output more than 30% faster. The 40% figure combines the lower rates with fewer tokens per task; that second part depends on your workload, which is why it is worth measuring rather than assuming.

The cheap cache reads matter most for long agent sessions that re-read the same prefix every turn. Prompt caching savings math shows how to work out what a hit rate is worth.

The benchmark table

Anthropic's announcement published scores for Opus 5.5, Fable 5.1 and Opus 5, in that order:

• Terminal-Bench 4.0 — 66.4% / 55.8% / 52.3%
• FrontierCode v1.1 (Main) — 54.4% / 50.3% / 48.0%
• CursorBench 4.0 — 57.8% / 51.8% / 46.6%
• GDPval-AA v2.1 — 1,846 / 1,735 / 1,708
• AutomationBench — 40.0% / 31.4% / 26.9%
• Humanity's Last Exam (with tools) — 67.7% / 65.6% / 63.6%
• OSWorld 2.1 — 81.8% / 80.7% / 74.0%

Two caveats from Anthropic itself. It says the gap between Opus 5.5 and Fable 5.1 is narrower than these scores suggest, and describes Opus 5.5 as performing at Fable 5.1's level on most work rather than above it. It also notes the model often suspects it is being evaluated, which complicates any evaluation, including the vendor's. Read the table as "comparable to Fable 5.1" rather than "better than Fable 5.1".

Effort now defaults to medium

This is the change most likely to surprise you without producing an error. Opus 5 and earlier Opus models default to high effort. Opus 5.5 defaults to medium. A request that omits effort now runs one level lower than it did.

Anthropic's documentation says effort names do not mean the same amount of thinking across models: in its testing, Opus 5.5 at medium exceeds Opus 5 at high on coding and knowledge-work evaluations, and low comes close on several coding tasks. It also warns that at a given level 5.5 tends to think more per turn than Opus 5, especially at xhigh and max. So copying your Opus 5 setting across is wrong in both directions. Set effort explicitly and run a fresh sweep.

Thinking also cannot be turned off. Effort is the only lever for latency and cost, and thinking tokens count against max_tokens even when their text is not returned. Size the limit for both.

Four breaking changes from Opus 5

1. Thinking cannot be disabled. {type: "disabled"} and budget_tokens both return a 400 at every effort level. Use low effort for fast routes.
2. Forced tool use is rejected. tool_choice of any or a named tool returns a 400. Use auto with strict: true and name the tool in the prompt, or structured outputs if you only wanted JSON back.
3. Thinking blocks are tied to the model and the conversation. Anthropic calls this preserved thinking. Harnesses that edit earlier turns need to become append-only; API accounts created on or after 31 August 2026 are enforced on this.
4. Computer use needs the new toolset on the Claude API and Google Cloud; the older tool returns a 400 there.

If any of these block you, Opus 5 remains available. The Opus 5 guide covers when staying put makes sense.

Opus 5.5 or Fable 5.1?

Fable 5.1 lists at $10 / $50 per million tokens, two and a half times Opus 5.5's rates. If Anthropic's "same level on most work" claim holds for your tasks, Opus 5.5 is the default and Fable 5.1 is the escalation tier for the hardest problems. The Fable 5.1 guide covers where the top tier still earns its price.

Safety behaviour differs from Opus 5. Opus 5.5 runs biology classifiers as well as cybersecurity ones, and most cybersecurity tasks are rerouted to Opus 4.8. Check stop_reason before reading content, and enable fallbacks. Anthropic also says it is available with zero data retention, and lists its retirement as not sooner than 22 September 2027.

How to check the cost claim

The catalogue lists the standard endpoint and a :batch variant, the same model on the asynchronous endpoint at half price. Neither changes the core question: is 5.5 cheaper per finished task on your work?

1. Take twenty completed tasks from your history, including some long agent runs.
2. Run each on Opus 5 at its current effort and on Opus 5.5 at low, medium and high.
3. Record output tokens, cache reads, turns to completion and whether you would accept the result.
4. Divide total spend by accepted results, per configuration.

If medium matches your Opus 5 baseline at fewer tokens, the 40% claim holds for you. How to benchmark LLMs on your own work has the harness.