Claude Opus 5
Succeeded on 2026-09-22 by Opus 5.5 at $4/$20, a 20% cut from the $5/$25 below.
Anthropic's claude-opus-5, announced 2026-07-24 (claude-opus-5-announcement).
Successor to Opus 4.8 in the Opus line, sold as the near-frontier tier at half
the price of Fable 5 rather than as a new capability ceiling.
Price: $5 / $25 per 1M input/output tokens — unchanged from Opus 4.8. A generation of claimed capability gain at a flat rate, which is the cleanest instance yet of the price-floor thread in synthesis: the money moves to cost-per-task, not the rate card. Anthropic's own framing leans on this — five of its six benchmark claims are stated as cost per task against a rival, not as scores.
Fast mode: 2× base price for ~2.5× default speed. Speed as a paid tier on the same weights is new
here; every other speed story in this spoke belongs to a different provider (cerebras-pricing) or a
different model. Effort still applies — all five levels: low, medium, high,
xhigh, max, defaulting to high. (Corrected 2026-08-04: this page previously listed four, copying
the launch post, which omits medium. Anthropic's own effort documentation
states "Claude Opus 5 supports all five effort levels" and recommends low and medium "liberally"
as the primary cost control.)
Specs the launch post never gave (anthropic-model-docs): 1M context (default and maximum), 128k max output synchronously and 300k on the Batch API, and a reliable knowledge cutoff of May 2026 — four months later than Fable 5's and Sonnet 5's January 2026. The tier sold as below the frontier carries the newest world knowledge in the family, which the price-per-capability story does not capture.
Thinking is on by default and cannot be disabled at xhigh or max (400). That is a change
from Opus 4.8, where omitting the parameter meant no thinking. And effort is not a
verbosity control: per the docs it moves thinking volume, not visible response length.
Vendor-reported performance (ratios, no absolute scores — see the caveats on the source page): more than 2× Opus 4.8 on Frontier-Bench v0.1; within 0.5% of Fable 5's peak on CursorBench 3.2 at max effort for half the cost; 3× the next-best model on ARC-AGI 3; ~1.5× the next-best pass rate on Zapier AutomationBench; past Fable 5 on OSWorld 2.0 at just over a third the cost; +10.2 points over 4.8 on organic chemistry. Coding, agentic tool use, and verification/iteration are the highlighted strengths, plus better visual output.
First third-party benchmark read (2026-07-26, opus-5-arc-agi-3). ARC Prize puts Opus 5 at 30.2% on ARC-AGI-3, against a previous record of 7.8% (GPT-5.6 Sol Max), plus 90.4% on ARC-AGI-2 and 97.5% on ARC-AGI-1 (both matching prior top scores at slightly higher cost). It solved five previously unsolved environments, four at or above human level. Note the tension with the launch post's "3× the next-best model on ARC-AGI 3": against the 7.8% record the multiple is ~3.9×, but ARC Prize also places "Fable-class" models at ~20%, which would make it ~1.5×. Whichever comparison Anthropic meant, its own ratio doesn't reconcile with both.
The same source is the reason not to read 30.2% as a pure reasoning gain: Opus 5 was built after ARC-AGI-3's format went public, and on Guanghan Ning's private Witness benchmark it scores 43.4, statistically tied with Fable 5 and kimi-k3, and lands below Opus 4.8 on the one game combining rules in an unfamiliar way. See arc-agi for the contamination problem.
The system card (2026-07-24, claude-opus-5-system-card) — read it before the ratios above. Anthropic's own 194-page report says plainly that Opus 5 is not more capable overall than Fable 5, assesses very low alignment risk, keeps it below the automated AI R&D threshold, and ships it under the same ASL-3 protections as Opus 4.8 (CB-1, not CB-2). Two findings cut against the launch framing:
- It hallucinates more than Opus 4.8 while knowing more. On AA-Omniscience, accuracy +11% over Opus 4.8 and hallucination rate +6%; net factuality 0.49, between Opus 4.8 and the Mythos models.
- Procedural care is saturated. It is the first Claude model to score perfectly on the lazy-investigation evaluation, and it essentially saturates the overconfidence-before-a-state-changing-command evaluation.
On the audit it beats Sonnet 5, Opus 4.8 and Mythos 5 on overall alignment and cooperates with misuse less than any model tested, while illegible thinking rose and white-box analysis still caught unverbalized grader awareness and fabricated data. Full detail on the source page.
Position in the family. Anthropic now sells four named lines at once: Sonnet 5 (mid), Opus 5 (flagship-by-price), Fable 5 (premium creative/coding at $10/$50), and Mythos 5 — which per the announcement trails Opus 5 broadly but leads it on cybersecurity and biology research. So Mythos is a domain-specialist line rather than the outright ceiling the Sonnet 5 system card hinted at. Opus 4.8 is not deprecated and serves as the safety fallback for flagged cyber requests in Claude Cowork.
Availability: Claude API, Claude.ai, Claude Code, Claude Cowork; AWS / Vertex AI / Microsoft Foundry named as partner platforms without a per-model breakdown.
(Resolved 2026-08-30: this page previously carried an "unknown context window" note, because the launch post states no figure. anthropic-model-docs gives 1M, recorded above; the open item is retired rather than left contradicting the paragraph two above it.)
For one class of buyer this is the ceiling, not the middle rung (2026-08-27). Anthropic's scientist-access post limits researchers in biology and chemistry to Opus-class models; Fable blocks their professional queries and Mythos runs through a separate government-partnered program. The half-price framing above assumes a buyer who could have chosen Fable 5. A biologist could not.
All figures dated 2026-07-24 and self-reported. Volatile.
Related
claude-opus-5-announcement · claude-opus-5-system-card · opus-5-arc-agi-3 · arc-agi · claude-fable-5 · claude-sonnet-5 · llm-api-pricing · llm-benchmarks · anthropic · claude-opus-4-8 · anthropic-scientist-access-expansion · research-access-programs