\n| Claude Fable 5<\/td>\n | $10<\/td>\n | $50<\/td>\n<\/tr>\n<\/table>\n That is the whole pitch in one table. You give up a slice of capability against Fable 5 and you halve your inference cost.<\/p>\n For a lot of workloads that trade is obvious. For some it is not, and we will get to which is which.<\/p>\n <\/span>What is actually new in Claude Opus 5?<\/span><\/h2>\n<\/span>The effort dial<\/span><\/h3>\nThis is the feature that deserves more attention than it got.<\/p>\n You can now tell the model how hard to think. Low, medium, high, and up through extra high and max on the API. Crank it up when the problem is genuinely hard. Dial it down when you just need a competent answer and you would rather not pay for a philosophy degree.<\/p>\n Anthropic says Opus 5 holds a surprising amount of its performance at lower effort levels while burning noticeably fewer tokens. On Zapier’s AutomationBench, even the lowest effort setting reportedly passes more tasks than any other model at any setting.<\/p>\n If you run agents at scale, this is the line item that matters. Model choice used to be your only cost lever. Now you have two.<\/p>\n <\/span>Fast mode<\/span><\/h3>\nSame model, roughly 2.5 times the speed, twice the base price. It is available on the Claude Platform and through usage credits in Claude Code.<\/p>\n Worth it for anything a human is sitting and waiting on. Probably not worth it for a batch job running overnight.<\/p>\n <\/span>Mid-conversation tool changes<\/span><\/h3>\nIn beta, and genuinely useful if you build agents. You can now swap which tools Claude has access to partway through a conversation without blowing up your prompt cache.<\/p>\n Anyone who has built a multi-stage agent knows the pain here. You either loaded every tool up front and watched the model get distracted, or you started a fresh conversation and paid to re-read your entire context. Now you can hand it a different toolkit at each stage and keep the cache warm.<\/p>\n <\/span>Automatic fallbacks<\/span><\/h3>\nAlso in beta. When a safety classifier flags a request, instead of the call just failing, it can route to a different model automatically.<\/p>\n In Claude.ai, Claude Code and Claude Cowork this happens by default, with flagged requests falling back to Opus 4.8. On the API you opt in.<\/p>\n This is a quiet but real reliability improvement. A hard block in the middle of an automated pipeline is an outage. A fallback is a slightly different answer.<\/p>\n <\/span>What do the Claude Opus 5 benchmarks say?<\/span><\/h2>\nAnthropic’s own numbers, so read them with the usual pinch of salt. That said, some of these gaps are wide enough to be interesting.<\/p>\n \n- Frontier-Bench v0.1:<\/strong> Opus 5 comes out ahead of every other model, and more than doubles what Opus 4.8 managed, at a lower cost per task.<\/li>\n
- CursorBench 3.2:<\/strong> at max effort it lands within half a percent of Fable 5’s best score, for half the cost per task.<\/li>\n
- ARC-AGI 3:<\/strong> this one tests novel problem solving, and Opus 5 scores three times the next-best model.<\/li>\n
- Zapier AutomationBench:<\/strong> around 1.5x the pass rate of the next-best model at the same cost per task.<\/li>\n
- OSWorld 2.0:<\/strong> a computer-use benchmark, where Opus 5 beats Fable 5’s best result at just over a third of the cost.<\/li>\n<\/ul>\n
Anthropic also calls it the new state of the art on GDPval-AA, a benchmark built to look more like real client deliverables than isolated coding puzzles.<\/p>\n One caveat worth stating plainly: independent head-to-heads against GPT and Gemini have not landed yet. The model is days old. Every number above is either from Anthropic or from an early-access partner, and partner quotes in a launch post are marketing, however honest they might be.<\/p>\n |