Claude Opus 5: Price, 1M Context, xhigh & Benchmarks (2026)

Anthropic released Claude Opus 5 on July 24, 2026 (official release / VentureBeat / MarkTechPost). Anthropic frames it as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price."

Four things that matter:

1. Large capability jump at unchanged pricing — $5 / $25 per M tokens (input / output), identical to Opus 4.8. Frontier-Bench v0.1 goes from Opus 4.8's 18.7% to 43.3% (max effort) / 44.4% (xhigh) — more than double
2. 1M context with 128K output (300K via Batch API) — 1M is both the default and the maximum. Minimum cacheable prompt drops from 1,024 to 512 tokens
3. A new xhigh effort level — above high, below the uncapped max. Aimed at long-horizon agentic and coding runs of 30+ minutes
4. Breaking changes — passing thinking: {"type":"disabled"} with xhigh or max returns a 400. Opus 4.8 is now legacy, and Opus 4.1 retires on August 5, 2026

This piece follows Claude Sonnet 5, Claude Opus 4.7, and the GPT-5.6 / Grok 4.5 API pricing comparison, and is based on day-one primary sources.

In Anthropic's published figures, the standout movement is not on classic coding benchmarks but on computer use, automation, and abstract reasoning.

The number with the most practical weight is OSWorld 2.0 going from 55.7% to 70.57%. For workloads that drive a browser, a terminal, or a GUI, that is the kind of gap that changes whether tasks which used to stall now complete. ARC-AGI-3 moving from 1.52% to 30.16% is more extreme still — an unusually wide jump for a single generation.

At the same time, note honestly that SWE-bench Pro lands at 79.2%, just short of Fable 5's 80.0%. Anthropic's own framing is that Opus 5 *comes close to* Fable 5: it replaces Opus 4.8, not Fable 5.

Pricing is identical to Opus 4.8.

- Input: $5.00 / 1M tokens
- Output: $25.00 / 1M tokens
- Fast mode: $10 / $50 (~2.5x faster)
- Up to 90% off with prompt caching, 50% with batch processing (as before)

The interesting part is the shape of the discount: the same per-token price now buys more than double the Frontier-Bench score. This is not a rate cut — it is an expansion of what a given budget can solve, which means cost-effectiveness only shows up if you measure total cost per completed task rather than per-token price.

The minimum cacheable prompt dropping from 1,024 to 512 tokens also has a direct effect on real spend for agent designs that fire many short system prompts.

Opus 5 adds xhigh to the reasoning-effort ladder, sitting above high and below the uncapped max.

- Intended for long-horizon agentic and coding work that runs 30+ minutes
- On Frontier-Bench, xhigh (44.4%) actually scores higher than max (43.3%) — max is not automatically optimal
- Thinking cannot be disabled at xhigh or max. Combining them with thinking: {"type":"disabled"} makes the API return a 400
- Thinking is on by default, with adaptive effort control adjusting depth as needed

Anthropic explicitly advises existing users to delete their verification prompts. Because the model now checks and verifies its own work, "always re-read and verify your answer" instructions written for earlier models become redundant and add noise.

The Opus 4.1 retirement on August 5 is deadline-bound work. Any codebase hard-coding old model strings is best audited now, right after launch. Anthropic has changed billing and access terms more than once during 2026 (the Agent SDK credit-billing change and its suspension, third-party tool restrictions on subscriptions), which strengthens the case for keeping model strings and billing assumptions in configuration rather than code.

The published figures also improve on the adversarial side.

- Gray Swan prompt injection: 2.0% attacker success rate (Opus 4.8: 5.5%)
- Browser-based attacks (Claude Cowork): 3.70% without safeguards, 0% in auto mode

If you are putting computer-use or browser-driving agents into production, the capability gain (OSWorld 70.57%) and the injection-resistance gain should be evaluated together. Permission design for that kind of rollout is covered in the Claude Cowork enterprise deployment guide.

- Opus 5 is now the default model on Claude Max, and the strongest option on Claude Pro
- Claude Code 2.1.212–2.1.219 ship Opus 5 support, alongside screen reader mode, MCP connector improvements, and stability fixes
- Available simultaneously on the API, Amazon Bedrock, Google Cloud, and Microsoft Foundry

For usage patterns see the complete Claude Code guide for 2026, and for running several agents in parallel, the Agent View deep dive.

The hierarchy has settled into three tiers: Fable 5 (frontier) → Opus 5 (production workhorse) → Sonnet 5 (high volume). The tier boundaries that blurred when Sonnet 5 out-benchmarked Opus 4.6 are redrawn here, with Opus repositioned as the agentic workhorse.

- Long-running coding agents (xhigh plus 30+ minute tasks)
- Computer use and browser automation — where the OSWorld 2.0 gain lands hardest
- Business-process automation workflows (Zapier AutomationBench 26.0%)
- Large-codebase analysis and long-document processing using the 1M context
- Multimodal work involving charts and screenshots (SWE-bench Multimodal 59.4%, Chartography 83.0%)
- High-volume generation needing 300K output via the Batch API

(1) Every number here is Anthropic's own. OSWorld 2.0 70.57%, ARC-AGI-3 30.16%, and Frontier-Bench 44.4% are day-one vendor figures. Convergence on independent leaderboards will take weeks.

(2) It is not a full Fable 5 replacement. SWE-bench Pro at 79.2% is short of Fable 5's 80.0%; the hardest workloads still warrant Fable 5 evaluation.

(3) max is not always best. On Frontier-Bench, xhigh (44.4%) beats max (43.3%). Effort levels need measuring on your own tasks.

(4) Thinking-on-by-default has a real cost. Thinking tokens bill as output. Even with unchanged rates, spend per request can rise — so compare total cost per task, not per-token price.

(5) There is dated migration work. Opus 4.1 retires on August 5, 2026.

Claude Opus 5, released July 24, 2026, replaces Opus 4.8 as the agentic workhorse while holding pricing at $5 / $25. The biggest movements are OSWorld 2.0 (55.7% → 70.57%), ARC-AGI-3 (1.52% → 30.16%), and Frontier-Bench (18.7% → 44.4%) — improvements weighted toward computer use and finishing long-horizon tasks rather than classic coding scores.

Three things to act on in code first: (1) find every place that pairs thinking: disabled with xhigh/max (it now 400s), (2) replace Opus 4.1 model strings before August 5, and (3) delete verification prompts written for older models to cut redundancy. After that, do not assume max — measure it against xhigh on your own workload.

Primary sources:
- Anthropic official release — Claude Opus 5
- Claude Opus product page

Coverage and analysis:
- VentureBeat — Anthropic launches Claude Opus 5
- MarkTechPost — Meet the New Claude Opus 5
- Axios — Anthropic releases new model, Opus 5