Claude Opus 5

anthropic/claude-opus-5 · by Anthropic · decoder-only-transformer

Pricing — 1 offering(s)

Anthropic

About this offering: Announced 2026-07-24. Fast mode (research preview) gives significantly faster output at $10 / $50 per 1M tokens, twice the standard rate, and isn't listed as a tier here (Anthropic pricing page, checked 2026-10-10).

Input tokens

$5.00 / 1M tokens (input) Current 2026-07-24 → present
Same as Opus 4.8 — no price change at launch.

Output tokens

$25.00 / 1M tokens (output) Current 2026-07-24 → present
Same as Opus 4.8 — no price change at launch.

Cache hits and refreshes

$0.50 / 1M tokens (cached input) Current 2026-07-24 → present
Cache reads (cache hits) cost 0.1x base input: $0.50 per 1M tokens.

5-minute cache writes

$6.25 / 1M tokens (cache write) Current 2026-07-24 → present

5-minute cache writes cost 1.25x base input ($6.25 per 1M tokens). 1-hour cache writes cost $10 per 1M…

5-minute cache writes cost 1.25x base input ($6.25 per 1M tokens). 1-hour cache writes cost $10 per 1M (Anthropic pricing page, checked 2026-10-10).

Batch input tokens

$2.50 / 1M tokens (input) Current 2026-07-24 → present
Message Batches price: 50% of the standard input rate. Fast mode isn't available with batches.

Batch output tokens

$12.50 / 1M tokens (output) Current 2026-07-24 → present
Message Batches price: 50% of the standard output rate.

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Claude Opus 5 fits into a cost-aware routing setup

About Claude Opus 5

Claude Opus 5 is the flagship model of Anthropic's Claude 5 family, announced on 2026-07-24. It is a decoder-only transformer trained with large-scale pretraining, instruction tuning and preference optimisation (Constitutional AI / RLHF), and supports an adaptive "thinking" mode controllable through an `effort` parameter with five levels (low / medium / high / xhigh / max), defaulting to high.

Anthropic publishes a 194-page system card for Opus 5 alongside the launch announcement. It is a capability- and safety-evaluation document rather than a product spec sheet, so some API parameters (such as the 128k maximum output tokens) are documented only in Anthropic's separate models-overview table. The 1M-token context window, by contrast, is confirmed both in that table and directly in the system card (Sections 8.9 and 8.10.2).

At launch Opus 5 became the default model on Claude Max and the strongest model available on Claude Pro. Anthropic reclassified the prior flagship, Opus 4.8, as a "legacy" model and recommends migrating, though 4.8 is not deprecated and remains available. API pricing is unchanged from Opus 4.8 at $5 / $25 per million tokens, with an optional "fast mode" at roughly 2.5x speed for 2x the base price.

Modelglass records this entry's capability confidence as low. Anthropic's launch materials give mostly relative performance claims (for example "more than doubles Opus 4.8's performance on Frontier-Bench v0.1") rather than published numeric scores, and the independent SWE-bench Verified figure (97%, Vals AI leaderboard, #1) is the main externally verifiable capability number. A GDPval-AA score, assumed in the originating task brief, was not found anywhere in Anthropic's Opus 5 materials and is not cited here.

Capability profile

How Claude Opus 5 rates across core capability dimensions, with the task-level evidence behind each rating.

reasoning Strong

Anthropic: 'more than doubles Opus 4.8's performance' on Frontier-Bench v0.1 (relative claim, exact score not published in the announcement or docs checked 2026-07-25 — see citations).

coding Strong

SWE-bench Verified (500 human-verified GitHub issues; the patch must pass the repo's own tests; no file-location hints): resolves 97%, so it fails fewer than 1 issue in 30. Independent (Vals AI leaderboard; Anthropic's own materials give only relative coding claims, no SWE-bench figure).

tool use Strong

Positioned by Anthropic 'for complex agentic coding and enterprise work'; Zapier AutomationBench 'around 1.5x the next-best model.'

instruction following Strong

context window Strong

1M-token context window confirmed against both the models-overview docs table (platform.claude.com) and the primary system card (194 pages, fetched in full via pdftotext 2026-07-25, bypassing the WebFetch 10MB ceiling) — Section 8.9/ProgramBench: 'Opus 5 episodes cover a range of context lengths up to the full 1M token window'; Section 8.10.2/BrowseComp: 'To extend beyond the 1M-token context window, we used context compaction.'

multilingual Strong

speed Moderate

Faster than Opus 4.8's 'weak' rating implies a step up: 'Fast mode' available at ~2.5x default speed for 2x base price ($10/$50 per MTok). Comparative latency listed as 'Moderate' in Anthropic's own docs table (vs Fable 5's 'Slower' and Sonnet 5's 'Fast').

cost efficiency Moderate

Anthropic markets it as 'near-Fable-5 intelligence at roughly half the price' — same $5/$25 pricing as Opus 4.8, but Anthropic's own benchmark framing (CursorBench 3.2, OSWorld 2.0) positions it as meaningfully more capable per dollar than the prior Opus generation.

Ratings are estimated — limited independent data is available for this model.

Coding benchmarks

View the full coding benchmark leaderboard →

Benchmark	Score	Harness	Source
swe-bench-verified ↗	97.0% 2026-07	—	📊 leaderboard ↗
terminal-bench-2-1	84.6% 2026-07	—	📊 leaderboard ↗
livecodebench ↗	89.0% 2026-07	—	📊 leaderboard ↗
terminal-bench-4-0	53.5% 2026-09	mini-swe-agent	📊 leaderboard ↗

Operator guidance

New default on Claude Max, strongest model on Claude Pro. Route here in place of Opus 4.8 for the most demanding reasoning/coding/agentic work. Opus 4.8 is not deprecated and remains available, but Anthropic now classifies it as a "legacy" model and recommends migrating. Drop to Sonnet 5 for the usual price/performance balance, or Haiku 4.5 for high-volume, latency-sensitive work.

Use cases

Complex agentic coding and enterprise work (Anthropic's own stated positioning)
Long-context analysis and multi-step tool/agent workflows
Tasks previously routed to Opus 4.8 where quality-per-dollar matters

Limitations

Benchmark claims in this doc are relative statements quoted from the launch announcement (e.g. '3x the next-best model'), not exact published scores — Anthropic did not publish numeric Frontier-Bench/CursorBench/ARC-AGI 3/Zapier AutomationBench/OSWorld 2.0 scores on the pages checked 2026-07-25
GDPval-AA was NOT found in the announcement, the system-card landing page, or the docs models-overview page — do not cite a GDPval-AA score for this model without a primary source
128k max output tokens is sourced from Anthropic's docs table only — the system card doesn't publish API parameter specs like max_tokens (it's a capability/safety-eval document, not a product spec sheet), so this is the correct source, not an unverified gap. Context window (1M), by contrast, IS cross-checked against the primary system card — see the context-window dimension above.