Claude Opus 5

Spec

Attribute	Value
Developer	Anthropic
Released	2026-07-24
Announced	2026-07-24
Context window	1M tokens
Pricing	$5/M input · $25/M output (standard); $10/M · $50/M (Fast mode, research preview)
License	proprietary
Availability	Claude.ai, API (claude-opus-5), Bedrock (anthropic.claude-opus-53), Vertex AI, Microsoft Foundry, Claude Code, Cowork
Additional specs:

• Max output (sync): 128k tokens
• Max output (Batch API + beta header): 300k tokens
• Knowledge cutoff: May 2026
• Tokenizer: new tokenizer (introduced with Opus 4.7); same text produces ~30% more tokens than pre-4.7 models

Release Date

July 24, 2026. Broadest simultaneous multi-platform launch in Anthropic's history: Claude API, Claude.ai (all paid tiers), Claude Code, Cowork, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry all available day-0.

Confirmed as the model behind the "Claude Honeycomb EAP" Cursor leak (July 8, 2026) — the xhigh effort level matches the "extra-high-effort mode" noted in that leak. The serial Fable 5 subscription extensions (through July 7, 12, 19) were Anthropic holding position while Opus 5 completed pre-release testing.

Benchmarks

Benchmark	Opus 5	Fable 5	Opus 4.8	Notes
Frontier-Bench v0.1 (agentic terminal coding)	43.3%	33.7%	~21.1%	Opus 5 leads by ~10pp
ARC-AGI-3 (novel problem solving)	30.2%	n/t	—	~4× better than GPT-5.6 Sol (7.8%)
SWE-bench Pro	79.2%	80.0%	69.2%	Within 0.8pp of Fable 5
SWE-bench Verified	96%	95%	—	#1 BenchLM leaderboard
GDPval-AA v2 (knowledge work Elo)	1,861	1,747	—	vs. GPT-5.6 Sol: 1,736
CursorBench 3.2 (max effort)	within 0.5% of Fable 5	—	—
BenchLM overall	85.88/100 (#1)	—	—	out of 215 models
Key headline: Frontier-Bench is the closest public proxy for real agentic engineering work. Opus 5 beats Fable 5 by ~10 percentage points at half the cost.

ARC-AGI-3 — the 30.2% is ARC Prize verified, on ARC Prize's standardized harness. On 2026-07-29 OpenAI reported GPT-5.6 Sol (and Terra, Luna) at 38.3% on the same public task set when run through the Responses API with retained reasoning and compaction enabled — against 7.8% for that model on the official harness. Nothing published states whether Opus 5 was measured with any equivalent state-preserving configuration, so 38.3% and 30.2% are not comparable and neither figure is evidence about the other model. See Eval Harness Configuration (source).

Leaderboards read 2026-08-16 — first place, and a gap smaller than its own error bar

LMArena's weekly snapshot puts the top three places with Anthropic, and the metric is the leaderboard's own percentage with a confidence interval — not Elo (source):

Rank	Model	Score	±	2026-08-09
1	Claude Opus 5 (High)	12.19%	±1.45%	11.99% (#1)
2	Claude Fable 5 (High)	12.01%	±2.57%	11.66% (#2)
3	Claude Opus 5 (Max)	11.95%	±1.71%	11.19% (#3)
The ordering is not a measured difference. #1 and #3 are 0.24pp apart
against intervals of ±1.45 and ±1.71; #2's interval is ±2.57 on its own. All three
overlap heavily, so "Opus 5 is ahead of Fable 5 on LMArena" is a statement the
snapshot does not support — only that the three are indistinguishable at the top.
This is the same discipline Eval Harness Configuration applies to a
figure without its harness: an interval that swallows the ranking is not a
ranking.

Two further movements in the same week, with no model released between the two snapshots:

• Every one of the top three rose (+0.20, +0.35, +0.76pp). A board that moves without a release is measuring its own vote mix, not a capability change.
• Claude Sonnet 5 left the top 10 — it was #9 at 7.46% ±2.15% on 08-09 and is absent on 08-16, with GPT 5.5 (High) at 7.73% taking the slot. Only the top 10 render server-side, so this records a departure from the visible table and not a score (source).
• LMArena renamed its own variant labels: Claude Opus 4.8 (Thinking) → Claude Opus 4.8 (High), same for Opus 4.7. Same models, relabelled by the publisher — recorded so a future diff of these snapshots does not read it as a new entry.

Artificial Analysis, same date, is a different measure and gives a different picture. Opus 5 holds the top Intelligence Index figure at 63, on both the max ($2.34 per task) and xhigh ($1.80) settings — unchanged from 2026-08-09 — with Claude Fable 5 (with fallback) at 62 for $3.14 and Opus 5 at high scoring 61 for $1.23 (source).

What changed around it is the price of parity. Grok 4.6 enters the same table at 61 for $0.84 per task — the same index as Opus 5 at high and as OpenAI's GPT-5.6 Sol at max, at roughly a third of Opus 5 max's measured cost. Artificial Analysis Intelligence Index is that publisher's composite of their own suite, not an accuracy, and Cost per Task USD is their measured cost of one task, not a per-token price.

These figures are prose rather than a table on purpose. claim-check.py compares a table row to its cited snapshot by (row label, column header) and strips parenthesised text from the label, so every Claude Opus 5 (…) effort variant collapses to one key and is compared against whichever row the snapshot lists last. A per-variant table here produced a false mismatch — 61 read against the (low) row's 52 — while every figure on the page was correct. Recorded in lint-2026-W33 (operator record, unpublished).

Key New Features

• Effort toggle: effort parameter — values low, high, xhigh. Defaults to high on the Claude API and Claude Code. Adaptive thinking is always on for Opus 5, Fable 5, and Opus 4.8 — no explicit thinking.type: "enabled" required.
• Fast mode (research preview): ~2.5× faster output at $10/$50 per MTok. Toggleable via /fast in Claude Code. Claude API only — not available on Batch API, Bedrock, or Vertex.
• Dynamic workflows (research preview): Claude Code plans a task, dispatches hundreds of parallel subagents with independent approaches, adversarial refutation, and iterative convergence.
• 128k output / 300k Batch: Largest standard output window in the Opus line.
• Default model on Claude Max; strongest model available on Claude Pro.

Use Cases

• Agentic coding and long-horizon software engineering tasks
• Novel problem-solving and multi-step reasoning
• Knowledge work and research at scale
• Computer use (autonomous desktop and browser control)
• Multidisciplinary reasoning (replaces Fable 5 as the recommended default for most enterprise/agentic workflows)

Compared To

vs. Claude Opus 4.8 ($5/$25 standard, SWE-bench Pro 69.2%)

Same price per token, same context window. Opus 5 delivers:

• +10pp SWE-bench Pro (79.2% vs. 69.2%)
• Roughly 2× Opus 4.8's Frontier-Bench score
• ~2× performance on knowledge work and software engineering (Anthropic claim)
• 128k vs. 32k max output

Opus 4.8 is now classified as a legacy model; Anthropic has published a migration guide.

vs. Claude Fable 5 ($10/$50 standard, SWE-bench Pro 80.0%)

Opus 5 costs half as much per token (standard tier) and:

• Beats Fable 5 on Frontier-Bench (43.3% vs. 33.7%)
• Beats Fable 5 on ARC-AGI-3 (30.2% vs. n/t)
• Beats Fable 5 on GDPval-AA v2 (1,861 vs. 1,747 Elo)
• Beats Fable 5 on SWE-bench Verified (96% vs. 95%)
• Trails Fable 5 on SWE-bench Pro by 0.8pp (79.2% vs. 80.0%)
• Within 0.5% of Fable 5 on CursorBench 3.2 at max effort

Fable 5 retains edge in highest-capability scenarios (cybersecurity: Mythos 5 leads); Opus 5 is positioned as the recommended default for most enterprise/agentic coding work.

Sources

• Introducing Claude Opus 5 — Anthropic
• Claude Opus 5 System Card (PDF)
• Pricing — Claude Platform Docs
• Models Overview — Claude Platform Docs
• Fast mode docs — Claude Platform
• AWS announcement
• Microsoft Foundry availability
• Bloomberg
• Claude Opus 5 on BenchLM
• → Anthropic announcement snapshot

Conflicting Reports

• Anthropic has published two figures for Fable 5's SWE-bench Pro, and this is not resolved. The Fable 5 launch announcement of 2026-06-09 gives 80.3% (source); the Opus 5 launch announcement of 2026-07-24 gives Fable 5 80.0% in its comparison table (source). Both are Anthropic's own announcements, so the schema's "official beats third-party" rule does not choose between them, and nothing in either says the benchmark was re-run. Each page therefore prints the figure from the announcement it is about, and both carry this note. Do not silently reconcile them — a later announcement restating an earlier number is not the same as a correction, and no source calls it one.