Anthropic platform model documentation

Anthropic's own developer documentation for the Claude model line — the operational reference, as distinct from the launch posts this spoke has been ingesting. Read across two pages: the models overview (specs, IDs, pricing, availability, cutoffs) and the effort reference (the levels and their per-model support).

T1, and a different kind of T1 from an announcement. Both are first-party, but a launch post is written to sell and a docs page is written to be depended on. Where they disagree, the docs win — and on Opus 5 they do disagree, in a way this spoke had faithfully reproduced.

The correction it forces

The launch post lists the effort levels as "low, high, xhigh, and max" — four. Both this spoke's Opus 5 pages copied that list. The effort documentation states it outright:

"Claude Opus 5 supports all five effort levels."

The missing one is medium, and it is not a footnote: the same page recommends using "low and medium liberally as your primary control for token cost and response time wherever your evals show quality holds." So the announcement dropped the level Anthropic's own guidance names as a primary cost lever.

Nothing suggests this was deliberate. But it is a clean instance of a pattern this corpus keeps finding from the other direction — llm-api-pricing-comparison pricing a retired DeepSeek model, secondary coverage widening claude-opus-4-8's "4× fewer unsupported claims" — here the vendor's own marketing is the lossy channel, and the fix was one page deeper on the same domain.

Specs the announcement never stated

The knowledge cutoff is the surprise. Opus 5's is May 2026; Fable 5's and Sonnet 5's are January 2026. The model Anthropic sells as below the frontier tier has four months more world knowledge than the tier above it — which the price-per-capability framing on claude-opus-5 does not capture, and which matters for any task where recency beats raw reasoning.

Availability is wider than the launch post says

The announcement lists Claude.ai, Claude Code, Claude Cowork, the API, and Claude Max. The docs add the enterprise channels: Amazon Bedrock (anthropic.claude-opus-5), Claude Platform on AWS (first-party IDs, not Bedrock-style), Google Cloud, and Microsoft Foundry.

Two API facts with product consequences

Thinking cannot be disabled at xhigh or max — those requests return a 400. Thinking is on by default on Opus 5, a change from Opus 4.8 where omitting the parameter meant no thinking at all.

Effort does not shorten answers. The docs are explicit that effort controls thinking volume, not visible response length — "changing effort does not reliably shorten responses, so prompt for length instead." That closes off the obvious lever for anyone trying to make Opus 5 terser, and it is the sort of thing only documentation says out loud.

Also worth carrying for cost work: changing effort between requests invalidates the prompt cache, because effort shapes the rendered prompt. Pick a level per workload, not per turn.

The pricing page, re-read 2026-08-30

A second docs page on the same domain — about-claude/pricing — and it carried a change the announcement channel never sent:

• Sonnet 5's introductory price is now the standard price. The $2/$10 per MTok rate was announced as introductory through 31 August 2026; the docs now say the scheduled rise to $3/$15 on 1 September will not occur. Corrected on claude-sonnet-5 and llm-api-pricing the day before the old deadline would have made both pages wrong.
• Fast mode is priced. Opus 5 and Opus 4.8 at speed: "fast" cost $10 / $50 per MTok — double the standard $5/$25, and exactly Fable 5's rate. Research preview, first-party API only, and it does not stack with the Batch API. So Anthropic now sells the same model at two prices, and the premium buys latency rather than capability — a second axis beside the effort parameter.
• US-only inference costs 1.1×. inference_geo: "us" on Claude 4.6 and later multiplies every token category. Data residency is a priced feature, not a policy toggle.
• Claude Managed Agents bills runtime as well as tokens — $0.08 per session-hour, metered only while a session is running, replacing container-hour billing.

The pattern from the Opus 5 correction above repeats: the docs hold facts with money attached that the launch channel does not carry, and the one that mattered here was a price that was scheduled to change and then wasn't.

The pricing page, re-read 2026-09-23: Opus 5.5

Read to check opus-and-sol-and-luna-willison. Opus 5.5 is listed at $4 / $20 per MTok (5-minute cache write $5, 1-hour $8, cache hit $0.20), batch $2/$10, fast mode $8/$40. The cache-hit multiplier is now per model: 0.05× on Opus 5.5, 0.025× on Fable 5.1 and Mythos 5.1, 0.1× elsewhere. Every other figure above is unchanged.

Why page the docs at all

This spoke's Anthropic coverage has been built almost entirely from launch posts and press. Those are the right source for positioning and the wrong one for facts a developer would act on — the announcement omitted the context window, the output ceiling, four deployment channels, the knowledge cutoff, and an effort level. Marked volatile: docs track the product, and this page is a dated read.

Related

claude-opus-5 · claude-opus-5-announcement · claude-fable-5 · claude-sonnet-5 · claude-opus-4-8 · effort-level · llm-api-pricing · llm-benchmarks · anthropic