Claude Opus 5 launches with auto-thinking, new caching rules, and different tool capabilities

Opus 5 runs at 5 and 25 per million tokens against Fable 5 at 10 and 50, so the top tier now costs double for a much smaller gap Thinking is on by default on Opus 5, which silently changes what a tight max_tokens setting does to your output Disabling thinking now returns an error above high effort, so any xhigh or max route that turns it off needs an audit before you migrate Prompt caching starts at 512 tokens on Opus 5, half the Opus 4.8 floor, so short reusable prompts cache with no code change In June I worked through whether Claude Fable 5 was worth double the price of Opus 4.8 and concluded that it usually was, for hard work. Claude Opus 5 landed on July 24 at Opus 4.8's exact price and closed most of that gap. So the answer changed, and a few of the changes will throw errors in code that worked last week. The Price Gap Held, the Capability Gap Closed Opus 5 costs 5 and 25 per million tokens, input and output. That is identical to Opus 4.8 and exactly half of Fable 5 at 10 and 50. Anthropic did not raise the sticker price on the tier it improved, which is the single most consequential fact in this release. What that buys, on the numbers: 79.2 percent on SWE-bench Pro against Fable 5's 80.3, and a CursorBench 3.2 result Anthropic describes as landing within 0.5 percent of Fable 5's peak at max effort, at half the cost per task. On OSWorld 2.0 it goes past Fable 5's best computer-use result at just over a third of the cost. A 1.1 point deficit on the headline coding row, for half the money. Last month the equivalent comparison had an 11 point spread. That is what actually changed, and it flips the default: Fable 5 used to be the reasonable choice for anything hard, and now it has to argue for itself on each task. There is a quieter cost lever too. The minimum cacheable prompt on Opus 5 is 512 tokens, down from 1024 on Opus 4.8. Prompts I had written off as too short to cache now create entries with no code change at all. If you run a lot of small repeated calls, re-check them before assuming nothing cached. Three Changes That Break Existing Code Migrating from Opus 4.8 is not a model string swap. Three things bite. Thinking is on by default. On Opus 4.8 and 4.7, omitting the thinking parameter meant the model did not think. On Opus 5, omitting it runs adaptive thinking. This is not just a behavior change, it is a truncation risk, because max_tokens is a hard cap on thinking plus response text together. Any route that never set thinking and sized max_tokens tightly around the expected answer can now run out of budget mid-response. Raise the ceiling or set thinking to disabled deliberately. Disabling thinking is capped at high effort. Setting thinking to disabled alongside an effort level of xhigh or max now returns a 400. Opus 4.8 accepted that combination happily. The check runs per request, so a later call that raises effort while thinking is still off fails even though earlier calls in the same conversation went through. Audit every call site, not just the first one. Rate limits are a separate bucket. Opus 4.8, 4.7, 4.6 and 4.5 share one combined pool. Opus 5 does not draw from it. Moving traffic across neither frees headroom on the old bucket nor inherits it, so check your tier's limits before shifting volume and assuming it will absorb. Two smaller ones worth knowing. Fast mode on Opus 5 runs at roughly 2.5 times default speed for twice the price, and it is a Claude API feature only, so it will not work on Bedrock, Google Cloud, or Microsoft Foundry. And Opus 5 ships with elevated cybersecurity safeguards that can decline a request outright, returning a normal 200 response with a refusal stop reason and an empty content array. Code that reads the first content block without checking the stop reason will break on that. Cyber-category refusals route to Opus 4.8 as the recommended fallback, so wiring up the server-side fallback parameter genuinely recovers the request rather than just relabelling the failure. The Behavior Shifts Nobody Puts in Release Notes This is the part that cost me the most time, because none of it throws an error. It just quietly makes output worse until you notice. Opus 5 verifies its own work without being told to. That means the instruction I had been carrying in prompts for two model generations, some variant of "include a final verification step" or "use a subagent to check this", now causes over-verification. The fix is to delete those lines rather than reword them. Anthropic's own migration guidance says the same, and it inverts the standard advice to ask a model to double-check itself. It also writes longer. Both user-facing responses and files it writes to disk run longer than on Opus 4.8, and lowering the effort level does not reliably shorten them. Prompting does. A short conciseness instruction is worth more here than any parameter. The subagent behavior reversed direction, which caught me out. Opus 4.8 under-reached for subagents and needed prompting to delegate. Opus 5 delegates readily, so the "delegate more" guidance I added for 4.8 is now actively expensive and wants a cap instead. Each subagent re-establishes context, explores, reports back, and then the coordinator reads the report, so the cost multiplies quickly on work that one agent could have finished in a few tool calls. It narrates its own corrections more than earlier models, flagging and explaining mistakes at length in a way that reads as thrash in anything user-facing. Scoping that down to corrections which actually change the reader's decision is a one-paragraph prompt fix, and worth doing before anyone else sees the output. Last one: it can expand scope, adding steps nobody asked for. A short instruction to deliver what was asked at the scope intended, and to flag disagreement in a sentence rather than silently redesigning the task, fixes most of it. One genuinely new capability sits alongside all this. Opus 5 supports changing a conversation's tool set between turns without invalidating the prompt cache, behind its own beta flag. Previously the tool list was fixed for a conversation's lifetime and any edit re-billed the entire prefix, which is why so many agent harnesses declare every tool up front and eat the context cost. If you have been working around that, it is worth a look. One caveat for anyone on guaranteed-latency plans: Priority Tier covers Opus 4.8 and Fable 5 but does not cover Opus 5, and a Priority Tier request naming it fails validation. Where Fable 5 Still Earns Double Fable 5 is not obsolete, it is narrower. It keeps the SWE-bench Pro crown at 80.3 and Anthropic still positions it for the most demanding long-horizon autonomous work, the overnight runs where a model has to stay coherent across hours without a human checking each step. It also has constraints Opus 5 does not. Thinking is always on and cannot be disabled at any effort level, so there is no cheap fast path. It requires 30-day data retention and is not available under zero data retention at all, which rules it out for some organisations before the capability question even comes up. So my routing rule now: Opus 5 by default for everything serious, including work I would have sent to Fable 5 in June. Fable 5 only for the longest autonomous runs, and only where the 1.1 point difference plausibly decides the outcome. Everything scoped and high volume goes down the ladder, on the same logic I used in Opus 4.8 vs Sonnet vs Haiku: How I Route Work in 2026. On effort, start at high (the default) and sweep down rather than reflexively reaching for xhigh. Low and medium are unusually strong on this model, and effort settings carried over from a previous model are rarely the right ones. Bottom Line The June answer to "is double the price worth it" was usually yes. The July answer is usually no, and that flipped without either model changing price. Opus 5 sits within about a point of Anthropic's top tier on the headline coding row at half the cost, caches shorter prompts, and runs a fast mode the top tier does not have. Budget an hour for the migration rather than a model string find-and-replace. The two changes that throw errors (thinking on by default eating your token budget, and disabled thinking above high effort) are quick to fix once you know they exist. The behavior shifts take longer, mostly because the fix is deleting prompt scaffolding that felt like good practice when you wrote it. I keep this kind of thing in the setup I package as Claude Blueprint, and I wrote the price side of the previous generation up in Claude Fable 5 vs Opus 4.8: Is Double the Price Worth It if you want the before picture.

Claude Opus 5 shipped on the 24th of July, 2026, as claude-opus-5. It's priced at $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8, and half of what Fable 5 costs. The context window is 1M tokens long, the max output is 128K. And there's one breaking change: the model thinks by default now. In other words, on the older Opus you had to ask for thinking explicitly, and it's now set up so it thinks on its own automatically. What actually changed from the previous Opus? Now it's allocating thinking depth per turn itself depending on the set effort level; the higher the effort, the more it thinks, but you can still set it explicitly in the request, and the parameter itself hasn't changed — thinking: {"type": "adaptive"} is still valid and still means the default. This means that when migrating a project that didn't do any thinking just by renaming it to this new Opus, you need to make sure you have enough room for both the thinking and the visible part of its response, so raise max_tokens. Also, if you disable thinking — which you can now do only at high effort or below, because pairing it with xhigh or max returns a 400 — it may output a call to a tool as plain text instead of emitting a proper tool_use block, or leak some of its internal tags into the output. So they recommend keeping it enabled and controlling the cost via the effort level instead. How much does it cost? It's priced at $5 per million input tokens and $25 per million output tokens, just like Opus 4.8. The price didn't move, but capability did. Fable 5 is $10 and $50 for the same volumes, so this is exactly half. There's also fast mode — a separate pricing line, and a research preview that's been running since February (fast-mode-2026-02-01), on Opus 4.8 as well. It's available only through the Claude API; it's not on Amazon Bedrock, Google Cloud or Microsoft Foundry. It's priced at $10 per million input tokens and $50 per million output tokens, so 2x the price for up to 2.5x the output tokens per second (for the record this is throughput, not time to first token, so don't expect end-to-end latency to halve). The minimum length a prompt needs to be to be cacheable is now 512 tokens instead of 1024, so prompts that were too short to get cached before are being cached automatically now. Should you use it or Fable 5? Opus 5 is half the price of Fable 5, and they're very close in terms of the numbers, so for most use cases it's usually about the price. According to their benchmarks: on CursorBench 3.2 at max effort it comes within half a percent of Fable 5 for half the price on OSWorld 2.0, which is computer-use work, it does even better than Fable 5 for a third of its cost it's state of the art on Frontier-Bench and GDPval-AA it more than doubles Opus 4.8 on Frontier-Bench v0.1 it scores 3 times higher than the next best model on ARC-AGI 3 it reaches around 1.5x the next best pass rate on Zapier's AutomationBench at the same cost These are Anthropic's figures, so take them with a grain of salt; they're grading themselves. But generally speaking — mid-tier is no longer a compromise pick, and if you were to use Fable 5 by default, it's worth reconsidering in light of these numbers. One place it doesn't win is exploit development. While on OSS-Fuzz it finds vulnerabilities about as well as Mythos 5, writing the exploits is scored separately — and there it's far behind. Which effort level should you use? The effort level starts at high by default; it's the default level both on the Claude API and for Claude Code, and this is what Anthropic document as a starting point. There are five levels in total — low, medium, high, xhigh and max — and in terms of this new Opus, it's more important than ever to choose the right one, as this model translates more effort into better output than any previous Opus, so it makes more of a difference which level you pick. We'd recommend testing yourself and adjusting up or down from high based on these results — lower when the quality is good enough to save time and money, higher for the hardest tasks. It's also worth noting that especially in this model, the two lowest levels are actually really capable, so the quality there is really good for a fraction of the cost, and the code review is very decent even at the lower levels. On the other hand, if you're planning to use the top two levels, make sure to set max_tokens high enough, 64K is a good place to start; you need it to have space for both thinking and acting across tools and subagents. Stream the responses rather than awaiting one of them. Why does it show a smaller context window than a million? The context window is 1M tokens long, and that's both its size and the max. It's not 200k long, this figure is about the client app; update it if possible. It's a single size, so you can't have reduced variants. What breaks when you switch? If you change the model to Opus 5, apart from renaming it, you need to update the surrounding configuration. The biggest potential pitfall is a prompt scaffolding that you might have left behind from older models — the model self-validates out of the box, so any leftover lines like "add a final verification step" or "use a subagent to verify" are prompting it to do additional verification now. Make sure to remove them. Two things simply aren't available on this Opus: web fetch and Priority Tier. If you're relying on either, that's a blocker rather than something you configure around. If you're an agent-infra maintainer, there are also 2 new features for you (both behind beta headers at the moment): you can now add/remove tools between conversation turns without losing the cached prompt, so you no longer need to resend a prompt with a fixed list of tools every time (beta header mid-conversation-tool-changes-2026-07-01) they have introduced a new fallback mode, which uses their recommended fallback models per refusal category instead of a list you maintain yourself (server-side-fallback-2026-07-01) — Claude API only, and explicitly not supported on Microsoft Foundry What does it do differently with no code change at all? If you change the model to this new Opus, you may notice a few differences in terms of behaviour even if you haven't changed anything in your prompts: longer answers and written deliverables talks more during agent sessions, narrating its steps is more eager to use subagents in multi-agent setups does self-verification None of these are bugs, they're all steerable by prompting. If you feel like the new Opus is slower or noisier than the previous one after the name change, make sure to check if it's not one of these differences before concluding that the model must've regressed; the "thoughtful and proactive" part of their description is the same thing from their side of the fence. In terms of its actual capabilities: sustains long chains of reasoning can implement features and do big refactors in agent coding across long tool loops, writing multi-file code without leaving stubs or placeholders behind does really good code review and finds a lot of real bugs with a low number of false positives is good at reading charts, documents, and diagrams builds multi-sheet spreadsheets and slide decks runs subagent teams without them stepping on each other's toes What we would do on day one In terms of the day 1 plan: rename the model check every max_tokens, as now it needs to accommodate thinking too; it's the most likely reason your output gets cut off after the switch find and remove any verification scaffolding in prompts and agent instructions make sure that anywhere you've disabled thinking, the adjacent effort level is high or below, as pairing it with xhigh or max will throw a 400; enforce this before shipping And perhaps most importantly: run your own code against it, not their benchmarks, and not ours either — although they could be a good start. We recently published a runnable benchmark with Kimi K3, Fable 5 and Opus 4.8 already hooked up. It contains a TypeScript ledger seeded with real bugs, with hidden tests and a prompt to benchmark the models against. Building with AI is our actual day job. We build software with AI for other companies — this is us reading the release notes on the tools we use every day. Curious how our benchmarks went? See how Kimi K3, Fable 5 and Opus 4.8 handled the same production task.