Claude guide: start on Opus 5.5, try Fable 5.2 before max

Anthropic’s new model guide on the Claude developer blog, posted November 9, makes Opus 5.5 at medium effort the starting point for most coding and knowledge work, and says to test Fable 5.2 before pushing Opus past high. If you already run Opus 5.5 at max on hard tasks, Fable 5.2 at high may get you the same result for less: on Anthropic’s FrontierCode 2.0 runs it scored 45.3% against Opus’s 45.5%, for roughly half the cost per task.

Where each model starts

Opus 5.5 is the default model in Claude Code and costs half as much per token as Fable 5.2: $4 / $20 per million input and output tokens, against $8 / $40. From there the guide moves down for narrower work. Sonnet 5.5, at $2 / $10, handles well-scoped coding such as a bug fix or a feature built against a spec, and starts at low or medium for chat and other replies a person waits on. A clear spec alone is not the test, though: a migration or a performance optimization can be fully specified and still call for Opus 5.5, or for Fable 5.2 if it is a major one.

Haiku 5.5 takes classification, routing, tagging and high-volume extraction, plus subagents that search files, read logs or summarize. Fable 5.2 is for the hardest projects in coding, finance and the sciences, for major refactors and architecture planning, for long unattended runs, and for cases where Opus 5.5 at high still falls short. Computer use and vision-heavy workflows start on Opus 5.5 at medium.

Defaults differ by surface. On the API, Haiku 5.5 and Opus 5.5 default to medium effort, while Sonnet 5.5 and Fable 5.2 default to high; Claude Code runs both of those at medium. /effort status shows what you are actually on.

Effort first, then a bigger model

The guide treats effort as the finer dial for both cost and capability. On SWE-bench Pro, Opus 5.5 at medium scored about 2.5 points below high for about 70% of the cost; at low it scored about 8 points lower for about a third of the cost; at xhigh it gained about 1.4 points for 2.5 times the cost of high. Past high, one more level can cost as much as a bigger model, so the advice is to test the next model up at a lower effort instead. Higher effort also adds latency, and it won’t give a model knowledge or skill it doesn’t have.

When a model isn’t good enough, the order is: make sure it can verify its own work with a test, a build or a script that calls the endpoint; raise effort one level; then move up a model, starting below the effort you used on the smaller one. If outcomes can be checked automatically, run everything at low and re-run only the failures at high. On the same SWE-bench Pro subset, Opus 5.5 run that way passed about 97% of tasks for about $0.17 each, against 95.3% for $0.29 with every task at high.

If the problem is speed rather than accuracy, Opus 5.5 and Fable 5.2 both offer fast mode on the Claude API, with higher output speed at premium pricing.

What drives the bill

The hardest tasks carry most of the spend, so the guide says to price them alongside the typical ones when comparing two models: on one 20-problem search benchmark, two problems accounted for 43% of it.

On long runs most input bills as cache reads, and Fable 5.2 and Opus 5.5 both charge $0.20 per million tokens for those. On a long, cache-heavy run the gap between them comes down to uncached input and output. Context size still counts on every turn: on Opus 5.5, a turn’s cache read costs about $0.03 at 150K tokens of context and about $0.004 at 20K. Hence the advice to keep the expensive model’s context for decisions and hand verbose reading to cheaper subagents, which bill at the rates of the model they run on; in Claude Code that means model: haiku or model: sonnet in the subagent’s definition.

Haiku 5.5 is the one model whose price depends on prompt length: $0.10 / $0.50 up to 100K tokens, $0.50 / $2.50 above. That upper tier is still a quarter of Sonnet 5.5’s per-token price, but when every call needs a long prompt the guide says to compare the two on cost per finished task, not list price.

Review is one of the repeated agent tasks the guide assigns to Sonnet 5.5 at medium. CodeRabbit, which moved simple and moderate reviews to Sonnet 5.5 in September, put its Claude cost per review at $0.47, against $1.16 on Sonnet 5.

Switching mid-project

The prompt cache belongs to the model that wrote it, so the first turn on a new model pays the cache-write price on the whole conversation. To shrink that turn, run /compact in Claude Code or start a fresh session with a short written plan; on the API, open the new model’s conversation with a summary instead of the full history.

Reasoning doesn’t always carry over either. Each thinking block records the model that produced it, and a model never reads the blocks of a model above it. When the new model can’t read them, the API drops them and the request still succeeds, so the work continues without that reasoning. A move from Fable 5.2 to any other generally available model drops its reasoning.