Anthropic's own flagship model, Fable 5, costs $10 per million input tokens and $50 per million output tokens. On July 24, 2026, the company released a cheaper model that beats Fable 5 outright on the benchmark Anthropic itself considers hardest: agentic coding. Claude Opus 5 launched at half that price, $5 and $25 per million tokens, the same rate as its predecessor Opus 4.8, and Anthropic's pitch for it was unusually direct: near-frontier intelligence, without frontier pricing.

This is Anthropic's fourth model release in under two months. Mythos 5, Fable 5, and Sonnet 5 all shipped in June 2026. Opus 5 arrives as what the company is positioning as the model most teams should actually run day to day, reserving Fable 5 for the hardest, longest-horizon problems.

The Benchmark Where the Cheaper Model Wins

The headline number comes from Frontier-Bench v0.1, an agentic terminal coding evaluation that measures whether a model can work through multi-step engineering tasks the way an autonomous coding agent would, start to finish. Opus 5 scored 43.3%, more than double Opus 4.8's 21.1% on the same test. That alone would be a solid generational jump. What makes it notable is what it beat along the way: GPT-5.6 Sol scored 34.4%, and Anthropic's own larger, more expensive Fable 5 scored 33.7%. On this specific test, the cheaper model outperformed the company's flagship.

Opus 5 also posted 30.2% on ARC-AGI-3, a benchmark specifically designed to resist memorization by presenting genuinely novel problems a model can't lean on training-data patterns to solve. That score is roughly three times higher than the next-best model tested, a gap wide enough that independent trackers flagged it as one of the more surprising individual results of the release. On GDPval-AA v2, a knowledge-work benchmark scored on a chess-style Elo rating, Opus 5 reached 1,861, and it posted 79.2% on SWE-bench Pro, a widely used software engineering benchmark.

What the Effort Toggle Actually Changes

Opus 5 ships with a new low, medium, and high effort setting, letting developers trade cost against capability on a per-request basis rather than committing to one fixed tradeoff for an entire application. That's a meaningful shift in how Anthropic is asking customers to think about model selection: instead of choosing between Opus and Fable tiers wholesale, a single Opus 5 deployment can flex its own computational effort depending on whether a given task is trivial or genuinely hard. A Fast mode is also available, running at roughly 2.5 times the speed for about twice the standard price, aimed at latency-sensitive applications where waiting on a slower, cheaper response isn't an option.

The model ships with a 1-million-token context window and a 128,000-token maximum output, alongside a knowledge cutoff of May 2026, the most current of any Claude model at launch. It's available immediately through the Claude API, Claude.ai, Claude Code, Claude Cowork, and through Amazon Bedrock, Google Cloud, and Microsoft Foundry, and it's now the default model on Claude Max and the strongest model offered on Claude Pro.

Anthropic's Safety Pitch This Time

Alongside the performance numbers, Anthropic emphasized Opus 5's alignment results, describing it as the most aligned model the company has released, with the lowest rate of deceptive behavior recorded in its automated safety audit to date. That framing matters given how much scrutiny frontier AI safety has drawn across the industry this year. Anthropic specifically noted that it deliberately avoided training Opus 5 on offensive cybersecurity tasks, even as the model's general capability improvements produced gains in that area as a side effect rather than a design goal.

That distinction is worth sitting with. Anthropic is drawing a clear line between capability that emerges naturally from a more generally intelligent model, and capability the company deliberately trained toward. Opus 5 still trails Mythos 5, Anthropic's specialized model, specifically on offensive cybersecurity benchmarks, a gap the company appears to be treating as intentional rather than a shortcoming to close in the next release.

Where Opus 5 Still Loses

Anthropic's own published comparisons don't claim a clean sweep, which is itself a notable choice for a launch announcement. GPT-5.6 Sol still edges Opus 5 on at least one agentic-coding evaluation, and on health and biology-specific tasks, the strongest results still belong to Mythos 5 rather than Opus 5. That's a more qualified pitch than most model launches make, and it suggests Anthropic is positioning Opus 5 specifically as the best price-to-performance option for everyday agentic and coding work, not as a universal replacement for every other model in its own lineup.

The Friction the Benchmarks Don't Capture

Not every reaction to the launch has been about the numbers. Developer community discussion following the release pointed to a persistent complaint that predates Opus 5 entirely: weekly and five-hour usage credit limits on Claude Max plans remain a real constraint for heavy users, regardless of how efficient the underlying model has gotten. A cheaper, more capable model doesn't automatically solve a rate-limiting problem if the plan structure around it stays the same, and that gap between model quality and account-level usage limits is shaping up to be the more practical friction point for developers actually trying to build on Opus 5 day to day, separate from anything the benchmark charts show.

Taken together, Opus 5's launch reflects where the frontier AI race has shifted this year: less about who posts the single highest number on one flashy benchmark, and more about who can deliver near-frontier results at a price and speed that makes deploying the model at scale, not just testing it in a demo, actually economical.