Claude Opus 5 launched July 24, 2026 at $5/$25 per million input/output tokens, exactly half the cost of Fable 5, while using one-seventh the reasoning tokens of its predecessor and matching its output quality.
Startups running Claude in production need to open a spreadsheet today. Anthropic dropped Claude Opus 5 on July 24, and the pricing math changes everything: $5 per million input tokens, $25 per million output tokens, half what Fable 5 costs at $10 and $50 respectively. That's not a modest discount. It's the kind of price drop that moves what's economically viable to automate.
The efficiency story underneath the price is just as striking. According to Niko Grupen, head of applied research at legal AI firm Harvey, Opus 5 matched the output quality of Opus 4.8 running at maximum reasoning while cutting average token usage by 26% overall. More specifically, it uses roughly one-seventh the reasoning tokens of Opus 4.8 to achieve the same result. Anthropic also ships the model with an adjustable effort toggle, letting engineering teams deliberately reduce reasoning depth on simpler prompts and save tokens where full intelligence isn't needed. For a startup processing millions of queries a month, that dial is worth real money.
The benchmark picture is harder to fit into the "near-frontier" framing Anthropic keeps using. On Frontier-Bench v0.1, which tests whether AI coding agents can complete real software engineering tasks end-to-end, Opus 5 scored 43.3%. Fable 5, Anthropic's own flagship, scored 33.7%. OpenAI's GPT-5.6 Sol hit 34.4%. On ARC-AGI-3, a test of novel problem-solving, Opus 5 scored 30.2% against GPT-5.6 Sol's 7.8%. These aren't "near" Fable 5 numbers. They're ahead of it, at half the price.
The money behind the model #
None of this is accidental. Anthropic raised $65 billion in May's Series H at a $965 billion valuation, making it the most valuable AI company in the world, ahead of OpenAI's $852 billion March valuation according to Korea JoongAng Daily. The round's strategic investors tell you exactly which problem Anthropic is solving: Samsung, SK Hynix, and Micron all participated as infrastructure partners. Three of the world's biggest memory chip manufacturers backing a single AI lab in the same round isn't a coincidence. It's a supply chain bet on sustained inference demand at scale.
What they're betting on is this: inference volume doesn't grow linearly with capability improvements, it grows when the cost per query drops to where new categories of use case become viable. Opus 5 is designed to widen that funnel. Running at less than half the latency of Opus 4.8, it's the new default on Claude Max and the strongest model available on Claude Pro, and it ships with a 1M-token context window with extended thinking on by default.
The OpenAI comparison is worth spelling out plainly. Opus 5 isn't competing with GPT-5 or whatever frontier model OpenAI is positioning as its best. It's competing with OpenAI's mid-tier: the models businesses actually run in production because the math works. At $5/$25 per million tokens with benchmark scores that beat Fable 5 outright, Anthropic has built something that makes that competitive calculus more difficult for OpenAI, not less.
What this means if you're building now #
There's a temptation to read Opus 5 as Anthropic telegraphing a retreat from the frontier race, leaving Fable 5 to hold the top line while they compete on price below it. Frankly, that reading doesn't hold up. A model that outscores your flagship on SWE-type benchmarks while running at half the cost isn't a concession. It's what happens when you get efficiency right. The 26% reduction in average token usage isn't just a cost story. It's evidence that Anthropic's training pipeline has gotten substantially better at producing useful output per token spent, which compounds across every workload running on the platform.
For founders and engineering teams building on Claude today, the question is simpler than the competitive analysis. Test Opus 5 this week. If you're currently running Fable 5 because you needed the capability, the benchmark data suggests you'll get better results on coding and agentic tasks at half the token cost - not marginally better, outright better. If you've been on Haiku or Sonnet because Fable 5 was too expensive, the same logic applies in the other direction. As VentureBeat noted, holding the price at Opus 4.8 levels while delivering these benchmark numbers is effectively a steep cut in the price per unit of capability. The production math changed yesterday. Also read: How Beijing became China's biggest tech venture capitalist and why the debate is just starting • OpenAI ships its first hardware product and it tells you where the coding tools war is headed • A developer ran a language model on an $8 chip and quietly broke the cloud AI model for IoT