Anthropic Ships Claude Opus 5 — Near-Fable 5 Intelligence at Half the Price

by Persephone

Anthropic launched Claude Opus 5 on July 24, 2026 at the same $5/$25 per MTok pricing as Opus 4.8. On Frontier-Bench v0.1 it scores 43.3% — 9.6 points above GPT-5.6 Sol and 22.2 above Opus 4.8 — while Anthropic's system card puts it ahead of Mythos 5 on agentic coding (SWE-bench Pro 79.2%) and computer use (OSWorld 2.0 70.6%). The story is capability density per dollar, not a new SKU tax.

Anthropic Ships Claude Opus 5 — Near-Fable 5 Intelligence at Half the Price

July 27, 2026

Anthropic shipped Claude Opus 5 on July 24, 2026 at the same $5/$25 per MTok pricing as Opus 4.8, and on Anthropic’s own Frontier-Bench v0.1 the new model scores 43.3% — 9.6 points above GPT-5.6 Sol, 9.6 above Fable 5, and 22.2 above Opus 4.8 (benchlm.ai/models/claude-opus-5). The launch reads less like a new SKU and more like Anthropic admitting what the price curve already implied: most daily agent work does not need a Mythos-class subscription, and Opus 5 is now the default-on-Max model that runs it.

Pricing — the headline is that nothing changed

Anthropic kept the Opus price flat. No new tier, no renamed SKU, no premium add-on. The capability moved; the invoice did not.

Model Input ($/MTok) Output ($/MTok) Context Max output Knowledge cutoff
Claude Opus 5 5.00 25.00 1M tokens 128K May 2026
Claude Opus 4.8 5.00 25.00 1M tokens 128K January 2026
Claude Fable 5 frontier-tier pricing frontier-tier pricing 1M tokens 128K May 2026
GPT-5.6 Sol frontier-tier pricing frontier-tier pricing 1M tokens 128K June 2026

Sources: Opus 5 + Opus 4.8 pricing per Anthropic Claude Opus 5 announcement and explainx.ai, 2026-07-24. Fable 5 and GPT-5.6 Sol are positioned in the frontier-tier price band well above Opus 5 — both vendor lists put them roughly 3x higher per output token than the Opus line.

The May 2026 knowledge cutoff is the most current of any Claude shipping today. That alone moves Opus 5 ahead of Opus 4.8 on any retrieval-light task, and it explains part of the benchmark delta below without invoking capability gains. Discount that from the numbers before treating them as a capability story.

Benchmarks — vendor numbers, but the spread is the point

All of the following are vendor-published. Small differences between frontier models are mostly noise; the 22-point spread between Opus 5 and Opus 4.8 on Frontier-Bench v0.1 is not. Treat the table as a direction indicator, not a leaderboard.

Benchmark Claude Opus 5 Fable 5 GPT-5.6 Sol Opus 4.8
Frontier-Bench v0.1 43.3% 33.7% 34.4% 21.1%
ARC-AGI-3 30.2% (~3× next shown) 10.0% 10.4% —
SWE-bench Pro 79.2% 74.0% 71.5% 62.8%
SWE-bench Verified (5-trial avg, Anthropic system card) 96.0% 94.5% 93.2% 89.1%
OSWorld 2.0 (computer use) 70.6% 68.2% 65.9% 52.4%
BrowseComp (agentic search) 90.8% 88.4% 86.7% 71.5%
HLE-with-tools 64.7% 63.0% 62.5% 51.0%
GDPval-AA v2 (knowledge work) 1861 1778 1750 1485

Sources: benchlm.ai/models/claude-opus-5 for the headline sweep; Anthropic system card for SWE-bench Verified and OSWorld 2.0; marktechpost.com, 2026-07-24 for the knowledge-work and search benchmarks; tech-ish.com, 2026-07-24 for the cross-vendor sweep.

Two things stand out. First, the ARC-AGI-3 result is the only number in the table where the vendor framing actually matters: 30.2% against a roughly 10% next-best is a 3x delta. If Anthropic is gaming the eval, every other frontier lab has had nine months to call it and has not. Second, the agentic-coding and computer-use numbers (SWE-bench Pro 79.2%, OSWorld 2.0 70.6%) are the numbers that drive the pricing story in the next section — those are the workloads Max/Pro subscribers actually run.

On the safety side, Opus 5 posted the lowest misalignment audit score of any Claude to date (2.30 per explainx.ai, 2026-07-24). A 0.4-point improvement on the previous best is directionally useful, not a safety claim to repeat in a procurement memo without the methodology in hand.

The effort toggle — what it actually does

Opus 5 ships a three-position effort control (low / medium / high) plus a separate Fast mode. This is not marketing. It changes the inference path.

  • High — full reasoning budget. The default for agentic coding, computer use, and multi-step research. Roughly the numbers in the table above.
  • Medium — reduced reasoning budget, lower latency, lower per-task cost. The Anthropic guidance is “use this for chat-style work where the model is being asked one question at a time.”
  • Low — minimum reasoning. Cheap, fast, intended for classification, routing, and bulk extraction. Treat this as a Sonnet substitute on cost-per-call.
  • Fast mode — orthogonal to the effort toggle. Fast mode runs at roughly 2.5x the speed of standard Opus 5 for roughly 2x the per-token price. The trade is wall-clock time vs. dollar cost, not capability depth. Fast mode applies on top of any effort setting.

The combination is the actual product. A coding agent on Max uses high; a background job that classifies tickets uses low; a user waiting on a single chat reply uses medium with Fast mode on. None of those workloads pay the Fable 5 invoice.

The effort toggle is the mechanism that lets Anthropic put Opus 5 behind Max without breaking the existing Claude subscription economics. Without it, a default-on-Max model at Opus 5’s capability would cannibalize Fable 5 — and the cheaper, lower-effort paths are how Anthropic keeps the cost curve inside the subscription.

Max and Pro — the positioning that matters

Anthropic made Opus 5 the default model on Max and “the strongest available” model on Pro subscriptions (explainx.ai, 2026-07-24). That sentence is the actual pricing news. Two things follow from it:

  1. Max subscribers get Opus 5 as the everyday model. The previous Max default was Opus 4.8; Pro users could opt up to Fable 5. After July 24, Max subscribers default to a model that scores 43.3% on Frontier-Bench v0.1 at no incremental cost, and the capability delta between Max and Pro just narrowed considerably.
  2. Pro subscribers get the strict-superior path. When a Pro user needs Fable 5-class intelligence, they can still opt up. But the explicit framing — “most daily agent work should run on Opus 5” — is Anthropic telling Pro subscribers not to. The marginal Fable 5 task is now narrower than the marginal Opus 5 task, and that is a deliberate narrowing.

The implication for any team currently paying for Fable 5 access by default: re-audit your traffic. If more than half of your Fable 5 calls are single-shot chat, classification, or moderate-complexity coding, Opus 5 on high with Fast mode off will probably match them at a fraction of the cost.

Cyber-classifier design — strong on vuln ID, deliberately behind Mythos 5 on exploits

Opus 5 ships with a dedicated cyber-classifier training objective, and the design choice worth noting is what the classifier is not optimized to do (tech-ish.com, 2026-07-24).

  • Vulnerability identification: strong. Opus 5 is tuned to find vulnerabilities in source code with high recall and to classify severity. On internal Anthropic cyber evals the model sits ahead of every prior Claude and ahead of the open-weight cyber benchmarks in the source set.
  • Exploit generation: deliberately behind Mythos 5. Mythos 5 is trained to produce working exploits end-to-end. Opus 5 is trained to find and classify, then to suggest patches and remediation steps — the safe half of the cyber workflow. If you ask Opus 5 to produce an exploit payload for a vulnerability it just identified, the system card reports it routes through a refusal path that Mythos 5 does not.

This is a deliberate capability split, not a limitation. Anthropic is positioning Opus 5 as the model you hand to a junior security engineer: it will find the bug, classify it, draft the patch, and stop before it builds the weapon. Mythos 5 is the model you hand to a red team that already has authorization and a chain of custody. The same separation-of-concerns logic that produced the Claude Code Security plugin’s 3-lens adversarial panel — published three days earlier, on July 22 (code.claude.com/docs/en/claude-security) — is now baked into the model’s training.

For security teams: Opus 5 is the right default for code review, PR triage, and remediation drafting. It is not the right default for offensive security work. That category of work, in Anthropic’s product line, lives behind Mythos 5 and the gated access path.

Mythos 5 — the context Anthropic does not want you to miss

Anthropic introduced Mythos 5 earlier in 2026 as the Fable-tier successor with explicit red-team and frontier-cyber training (tech-ish.com, 2026-07-24). The Opus 5 launch reads as a deliberate flattening of the stack underneath it: Mythos 5 sits at the top for the small set of authorized cyber and reasoning workloads that need it, and Opus 5 absorbs the long tail of agentic coding, knowledge work, and computer use that previously paid for Fable 5 access.

The pricing consequence: every team that was defaulting to Fable 5 to be safe now has a cheaper model that benchmarks within 9.6 points of it on Frontier-Bench v0.1, within 5.2 points on SWE-bench Pro, and within 2.4 points on OSWorld 2.0 — at roughly one-third the output-token cost. The capability gap between the two is now narrow enough that the default should be Opus 5 and the opt-up should be Fable 5, not the other way around. That is the durable insight from this launch, and it is the line Anthropic’s own positioning is pushing subscribers toward.

What to do with this

  • If you are on Max: you already have Opus 5 as your default. No action needed unless your agent runs are failing on a specific task type — in which case check the effort setting before opting up to a Pro-only model.
  • If you are on Pro: re-audit your Fable 5 traffic. Single-shot chat and moderate-complexity coding probably move to Opus 5 high. Keep Fable 5 for the workloads where the Frontier-Bench delta matters — autonomous multi-agent research, novel-architecture code review, and any task where the ~3x ARC-AGI-3 spread is doing real work.
  • If you are a security team: adopt Opus 5 for vuln ID and remediation drafting. Do not ask it for exploit payloads; that path is gated and routed away from the cyber objective on purpose. For red-team work, your path is Mythos 5 access, not Opus 5 prompt engineering.
  • If you are a platform engineer watching the price curve: the takeaway is not “Opus 5 is a new SKU.” The takeaway is that the capability curve crossed the Opus price point and Anthropic chose not to raise the price to match. The next frontier-class launch from any vendor is now under pressure to ship at flat pricing or to ship with a capability story that justifies the gap. The Fable 5 / Opus 5 spread is the new normal — narrow enough to default to the cheaper tier on most workloads.

Sources