Claude Opus 5

Why this oneWhat it is actually for

This is the model Anthropic tells you to start with — the overview page opens with "start with Claude Opus 5 for complex agentic coding and enterprise work." You pick it over Sonnet 5 when the task is hard enough that a quality gap costs more than the price gap, and over Fable 5 when you have not measured a gain worth double the rate. Two properties are specific enough to decide on. Its reliable knowledge cutoff is May 2026, the latest of any Claude model, which matters if you are asking about recent libraries or events. And it is the only model besides Opus 4.8 with fast mode, giving up to 2.5x higher output tokens per second when latency is the constraint.

What it isArchitecture, lineage, training

Opus 5 is a closed-weight model available only through APIs. Anthropic discloses no architecture: no parameter count, no dense-versus-mixture-of-experts statement, no expert count, no layer, head, or vocabulary numbers, and no description of the training corpus or compute. There are no released weights, so there is no Hugging Face model card and no config.json to read architecture numbers from. Treat any published parameter figure for this model as unsourced.

What is documented is the serving contract. The API model ID is claude-opus-5 — from the 4.6 generation onward, model IDs "use a dateless format that is also a pinned snapshot, not an evergreen pointer," so this string will not silently change under you. The context window is 1M tokens and max output is 128k on the synchronous Messages API. Reliable knowledge cutoff and training data cutoff are both listed as May 2026.

Its position in the lineage is defined by two behavioural changes from Opus 4.8 rather than by any published architectural difference. First, thinking flipped from off-by-default to on-by-default: the per-model table lists Opus 4.8 as default "Off" and Opus 5 as "On." Second, the prompt-cache minimum halved from 1,024 tokens to 512, so prompts previously too short to cache now cache with no code change.

It shares the newer tokenizer introduced with Opus 4.7, which the pricing page says "produces approximately 30% more tokens for the same text" than the tokenizer used by Sonnet 4.6 and earlier. Its list price is identical to Opus 4.8's, so the generation change brought capability rather than a price move.

At a glanceSee it

An Opus 5 request, showing the effort dial and the gate that blocks disabling thinking above high effort.

CapacityContext, output and what fits

The parametersEvery knob, and what moving it does

Three knobs carry the weight. output_config.effort is the real cost dial, and the documented advice is unusual: sweep downward, because low and medium hold quality on many workloads. max_tokens needs revisiting on any route migrated from Opus 4.8, since thinking now runs by default and eats the same budget as the answer. And thinking deserves a deliberate decision rather than a carried-over default — disabling it is what triggers this model's documented habit of writing tool calls into plain text where they never execute.

SamplingShaping the output distribution

Opus 5 has no usable sampling surface. The thinking documentation states that on Opus 5, "non-default temperature, top_p, or top_k values return a 400 error on every request, regardless of whether thinking is used," and the deprecation page lists all three as deprecated from Opus 4.7 onward with the replacement being to omit them and use prompting. There is no seed, so there is no determinism knob either — and the API reference is explicit that even at temperature zero on models that accept it, "the results will not be fully deterministic."

The practical consequence is that the output distribution is shaped by effort and by your prompt, nothing else. A useful and slightly counterintuitive detail: the effort docs note that on Opus 5, "changing effort does not reliably shorten responses" — effort controls thinking volume, not visible verbosity. If you want shorter answers, ask for shorter answers. The sensible starting point is to send no sampling parameters, leave effort at the high default, then run an effort sweep on your own evaluation set rather than reusing settings from an earlier model.

ReasoningThinking, effort and budgets

Thinking is on by default. The per-model configuration table lists Opus 5 as "Adaptive only" with default "On" — omitting the thinking parameter runs adaptive thinking, which is a change from Opus 4.8 and 4.7 where omitting it meant no thinking. Depth is controlled by output_config.effort; budget_tokens is gone and returns a 400.

You can turn thinking off, but only partway: thinking: {"type": "disabled"} is accepted at effort high or below, and returns a 400 at xhigh or max. The check runs on every request, so a later call that raises effort while thinking is still disabled fails even though earlier calls in the same conversation succeeded.

Thinking is billed as output tokens whether or not you see the text, and counts toward max_tokens. The raw chain of thought is never returned; display defaults to "omitted" and "summarized" yields a summary written by a different model. Disabling thinking has a documented side effect: the model can emit tool calls as plain text, which never run.

ToolsFunction calling and server tools

Standard function calling via tools, with tool_choice of auto, any, tool or none, and disable_parallel_tool_use available on each. Forced tool choice works here because the restriction that blocks it applies only to manual extended thinking, which this model rejects. Parallel calls are on by default — return all tool_result blocks in one user message, since splitting them across messages trains the model out of parallel calling.

Tool definitions cost tokens: the pricing page puts the tool-use system prompt at 286 tokens for auto/none and 406 for any/tool, the lowest overhead of any current model. Structured output uses output_config.format with a JSON schema, and strict: true guarantees tool inputs validate.

The failure mode worth designing around is disabling thinking: the troubleshooting page reports the model "occasionally writes a tool call into its text instead of emitting a tool_use block," which never executes and silently pollutes later turns. Leave thinking on and lower effort instead.

CostPrice, caching, batching, what drives the bill

List price is $5 per million input tokens and $25 per million output tokens — the same rate as Opus 4.8, and half of Fable 5. This is not a price cut; the Opus tier has held this rate, and what changed is which model occupies it.

Four things drive the real bill. Caching is unusually favourable: the 512-token minimum is the lowest of any Claude model, so short prompts that never cached on Opus 4.8 now do. Thinking is on by default and billed at the output rate, so a route migrated from a no-thinking Opus 4.8 setup gets more expensive with no code change. The tokenizer emits roughly 30% more tokens for the same text than Sonnet 4.6 and earlier, so cross-generation per-token comparisons mislead. And inference_geo: "us" adds a 1.1x multiplier across every category.

Where it runsSurfaces and availability

StrengthsWhat it is good at

• Anthropic's own stated default — the models overview says to start here for "complex agentic coding and enterprise work."
• May 2026 reliable knowledge cutoff, the latest of any current Claude model and four months ahead of Fable 5 and Sonnet 5.
• 512-token minimum cacheable prefix, half that of Opus 4.8 and Sonnet 5, so short system prompts cache with no code change.
• Lowest documented tool-use system-prompt overhead of any current model at 286 tokens for tool_choice: auto.
• Fast mode delivers up to 2.5x higher output tokens per second on the same weights, with usage.speed reporting which speed actually served the request.

LimitsWhere it falls down

• Thinking is on by default, so a route migrated from Opus 4.8 silently gains thinking tokens billed at $25/MTok and can truncate against an unchanged max_tokens.
• Disabling thinking is capped at high effort — combining it with xhigh or max returns a 400, validated on every individual request.
• With thinking disabled the model can write tool calls into plain text, which never execute and leave no error, and can leak internal XML tags into visible output.
• No sampling controls at all: temperature, top_p and top_k all 400 on non-default values, and there is no seed.
• Fast mode is a research preview requiring account-manager access, and is unavailable on every third-party platform and with the Batch API.

Against its neighboursHow it compares

Against Claude Sonnet 5, the gap is price and knowledge, not context: both have 1M windows and 128k max output. Opus 5 costs $5/$25 against Sonnet 5's $2/$10, and has a May 2026 cutoff against Sonnet 5's January 2026. Sonnet 5 is the sensible production default; Opus 5 is what you escalate to. Against Claude Opus 4.8, the list price is identical, so the comparison is purely behavioural: Opus 5 has the later cutoff, the halved cache minimum, thinking on by default, and lower tool-use overhead. Against Claude Fable 5, Opus 5 is half the price, has a later cutoff, supports fast mode, and lets you disable thinking — Fable 5 buys the higher capability tier plus a refusal path you must handle.

Getting startedThe smallest call that works

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "max_tokens": 4096,
    "messages": [
      {"role": "user", "content": "Explain what a write-ahead log is."}
    ]
  }'

Change max_tokens first. Thinking runs by default here and shares that ceiling with the answer, so a value tuned on a no-thinking model will truncate. Then run an effort sweep — add "output_config": {"effort": "medium"} and compare against the high default on your own cases, since Anthropic recommends using the lower levels as the primary cost control.

SourcesWhere every claim above came from

• Models overview — Anthropic
• Pricing — Anthropic
• Effort — Anthropic
• Thinking — Anthropic
• Troubleshooting thinking — Anthropic
• Fast mode (research preview) — Anthropic
• Prompt caching — Anthropic
• Structured outputs — Anthropic
• Model deprecations — Anthropic
• Messages API reference — Anthropic
• Could not be confirmed from a primary source as of 2026-07-25: every architecture detail — parameter count, dense versus mixture-of-experts, expert count, layers, heads, vocabulary, or training compute. Anthropic publishes none for Opus 5 and releases no weights, so no model card or config.json exists. No benchmark scores appear on this page because none were verified against a primary source in this session. The exact release date is not stated in the docs read here; the deprecation table's "not sooner than July 24, 2027" retirement date is the only dated anchor. Whether Opus 5 carries the safety classifiers documented for Fable 5 could not be confirmed — the Fable 5 launch page attributes them to Fable 5 specifically, and no page read here extends that to Opus 5.

RelatedGo deeper