Claude 5 Opus Is Here, and It Changes the Math

Anthropic just released Claude 5 Opus, and it lands in an interesting spot below the flagship Fable class but above the workhorse Sonnet series. It is aimed at a middle ground where it needs to be powerful enough to matter while staying cheap enough to use constantly without worrying about the bill.

The pricing is simple. The standard variant costs five dollars per million input tokens and twenty five dollars per million output tokens. There is also a fast mode that doubles the speed and doubles the price along with it, landing at ten dollars per million input tokens and fifty dollars per million output tokens. That is a clean tradeoff without hidden tiers or confusing rate limits, so if you are running latency sensitive agent workflows the fast mode probably pays for itself, while batch processing overnight probably does not need it.

Released on July 24, 2026, Opus 5 keeps the same $5/$25 token price as Opus 4.8 and Opus 4.7 before it, with a 1M-token context window and that Fast Mode option on top. For orientation, the rest of the current Anthropic ladder looks like this: Sonnet 5 at $2/$10 is the fast general-purpose tier, Haiku 4.5 at $1/$5 handles high-volume simple work like classification and extraction, and Fable 5 at $10/$50 is the flagship. (Mythos 5, Fable's architecture sibling, remains a $10/$50 preview.)

There was even some drama getting here: in early July, the US Department of Commerce lifted export restrictions on Fable 5 and Mythos 5, and Anthropic restored global access to both. If you noticed flagship-tier Anthropic models flickering in and out of availability across regions this summer, that was why.

It is worth mentioning that Opus 5 ships with lower safety guardrails than Fable. Anthropic seems to be treating the mid tier model differently than its flagship, presumably because Opus is not being marketed as the safe default the way Fable is. That will be a plus for some use cases and a dealbreaker for others depending on what you are building.

The part that actually caught me off guard is the benchmark performance. Opus 5 scores sixty one on the Artificial Analysis leaderboard, which does not just beat the Sonnet line but beats Fable 5 outright, making Opus 5 the highest scoring model on the entire leaderboard right now. That is a strange result for what is supposed to be the middle model in the lineup. On top of that, Artificial Analysis also ranks it as the best model for AI agents, not just the best in its price range but the best overall.

It helps to know what the score measures before deciding how much a leaderboard win matters. The current Artificial Analysis Intelligence Index aggregates nine evaluations into a single 0–100 number: GDPval-AA v2 (agentic tasks across 44 real occupations, graded by blind expert comparison), 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. Those roll up into four categories — agents, coding, general capability, and scientific reasoning — weighted 25% each.

The detail that makes the leaderboard credible is that Artificial Analysis runs every component benchmark itself rather than accepting vendor-reported numbers. That is exactly why a mid-tier model topping the index is such a big deal: an independently-run composite that is a quarter agentic evaluations is not the kind of test you can ace with a launch-day marketing chart. Anthropic's own positioning has already shifted to match — Opus 5 is now described as the recommended starting point for complex agentic coding and enterprise work.

A benchmark win on its own is nice, but a benchmark win at half the price of the model it just beat is a much bigger deal. Opus 5 is incredibly close to Fable in practice, close enough that most workloads probably will not notice the difference, while costing half as much. That kind of gap changes how teams think about their default model choice, because when the cheaper option outperforms the flagship on the metric that matters most and costs half as much, the flagship becomes a much harder sell.

Per-token prices hide the real costs, so here is the worked version. Take a long agentic coding run that consumes 120K input tokens and 450K output tokens — a realistic overnight agent job: - Opus 5: (0.12 × $5) + (0.45 × $25) = $11.85 - Fable 5: (0.12 × $10) + (0.45 × $50) = $23.70 Same job, half the bill, on the model that benchmarked better.

Now layer on the discounts most people forget. Anthropic's prompt caching charges only 10% of the input rate for cache hits, which on Opus 5 is $0.50 per million tokens — so an agent that re-reads the same stable 100K-token context ten times pays about $0.50 for those re-reads instead of $5.00. The Batch API halves both input and output prices (an effective $2.50/$12.50 on Opus 5) for anything that can run asynchronously overnight. And Fast Mode doubling the price is really a question of whether halving wall-clock time saves more than it costs — for interactive agent loops it usually does; for batch jobs it never will.

For context outside the Anthropic tent: GPT-5.6 Sol runs $5/$30, so Opus 5 matches its input price and undercuts its output by about 17%, while Gemini 3.1 Pro runs $2/$12 on prompts up to 200K tokens. Opus 5 at $5/$25 lands right at the sweet spot of that table. And if you prefer subscriptions to API math, Opus 5 is also included in the Claude Pro ($20/month) and Max ($200/month) consumer plans.

Claude 5 Opus ends up being one of those releases where the positioning undersells the product. It is marketed as the middle child, but right now it sits at the top of the leaderboard, ranks as the best agent model available, and costs dramatically less than the model it just dethroned. The looser guardrails are a real consideration depending on your use case, but in terms of capability per dollar, Opus 5 is hard to argue with. If you have been defaulting to Fable out of habit, it is worth running your workload through Opus 5 first, because you might not go back.