# Claude Opus 5 Claude Opus 5 is [[Anthropic]]'s Opus-tier [[Claude]] model, released 24 July 2026. API string `claude-opus-5`. See [[Claude 5]] for how the whole generation fits together. The short version: it lands very close to [[Claude Fable 5]] on most work, at half the price, and it is the first model Anthropic is willing to describe as genuinely hard to prompt inject. ## Pricing | | Input | Output | |---|---|---| | Standard | $5.00 | $25.00 | | Fast mode | 2× base | 2× base | Unchanged from Claude Opus 4.8. Fast mode buys roughly 2.5× default speed for 2× the price, which is a rare case of a speed premium that is close to honest. ## Benchmarks Anthropic-published, so treat as directional: | Evaluation | Result | |---|---| | Frontier-Bench v0.1 | State of the art. More than double Opus 4.8's score, at lower cost per task | | CursorBench 3.2 | Within 0.5% of [[Claude Fable 5]], at half the cost | | ARC-AGI 3 | Three times the next-best model | | Zapier AutomationBench | ~1.5× the next-best pass rate | | OSWorld 2.0 | Beats Fable 5's best result at just over a third of the cost | | Organic chemistry (internal) | +10.2 points over Opus 4.8 | The CursorBench and OSWorld numbers are the ones that matter commercially. If you are within half a percent of the frontier model at half the price, the frontier model becomes a specialty purchase. ## What's new - **Self-verification and longer horizons.** It checks its own work and keeps going rather than stopping at plausible - **Autonomous tool creation.** Denied direct image access during testing, it built its own computer vision pipeline to get the job done. It also builds test harnesses unprompted - **Mid-conversation tool changes** (beta) - **Automatic fallbacks to alternate models** (beta) - **Better visual output** - Notably stronger on organic chemistry and protein research ## The security story This is the part [[Boris Cherny]] singled out as more interesting than any eval score: > Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully. Layer the defenses (model alignment, plus prompt injection probes, plus Auto Mode in [[Claude Code]]) and he reports the attack success rate dropping to approximately zero. Take the number with the usual caution, since it comes from the vendor. But prompt injection has been the unsolved structural problem in agentic AI for two years, and "we layered three defenses and got to ~0" is a more useful claim than another benchmark point. See [[Claude Code Auto Mode]]. On the safety side, Anthropic calls it their most aligned model to date with the lowest misalignment scores. Cybersecurity classifiers intervene about 85% less often than on Fable 5 while keeping vulnerability discovery ability, which mostly says the false-positive rate on Fable was rough. ## Reasoning effort: more is not better Vals AI ran Opus 5 across all five reasoning levels on Vibe Code Bench. Performance climbs from low to medium to high, then **gets worse at the two highest levels** while costing significantly more. The regression sits within error bars, so read it as "no gain," not "actively harmful." The practical rule: **high is the ceiling worth paying for.** This matches what [[GPT-5.6]]'s own migration guide says about testing one level lower than your current default. Two labs, same message in the same month. The reflex to crank effort to maximum is now measurably wrong on both frontiers. ## Reception The [[Hacker News]] thread hit 1,777 points and 1,329 comments, which is enormous even for a model launch. - Genuine amazement at the computer-vision-pipeline story. One comment captured it: less than four years since ChatGPT, and models now spontaneously build their own ML pipelines - The sharpest criticism was about **token burning**. If the model hits a wall, it should ask before spending your budget on an elaborate workaround. Autonomy that does not check in is a billing risk dressed as capability - Recurring skepticism about whether it invented anything or recombined training data - A theme worth noting for anyone building with it: working alongside these models **increases cognitive load** even while raising output, because context management and code review never stop Every's hands-on verdict was cool: "a hard model to love." In its first week it argued with instructions, stopped before finishing work, and did not slot into their existing plugin setup. Their framing is that it can do impressive work but may require dismantling the systems you already built around older models. That last point is not a complaint about the model. It is the direct consequence of the context engineering shift below, and it is the reason this release is more disruptive than the version number suggests. ## The context engineering reset Anthropic removed **over 80% of Claude Code's system prompt** for Opus 5 and Fable 5, with no measurable loss on their coding evals. If your CLAUDE.md, skills, and prompts were tuned for older models, they are now probably fighting the model rather than helping it. Run `/doctor` in [[Claude Code]]. The full set of changes is documented in [[Context Engineering]]. ## Caveats - All benchmark figures are Anthropic-published. No independent replication yet - Mid-conversation tool changes and automatic model fallbacks are both beta - The prompt injection claim comes from the vendor's own system card and red teaming - Every's experience suggests real migration friction if you have an elaborate harness ## References - Announcement: https://www.anthropic.com/news/claude-opus-5 - System card: https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf - [[Boris Cherny]] on prompt injection resistance (2026-07-24): https://x.com/bcherny/status/2080713091688583312 - Vals AI on reasoning levels (2026-07-25): https://x.com/ValsAI/status/2080817011157094511 - [[Hacker News]] discussion (1,777 points): https://news.ycombinator.com/item?id=49038433 - Every, "Vibe check: Opus 5" (paywalled): https://every.to/vibe-check/opus-5 - System prompt leak by Pliny (2026-07-25): https://x.com/elder_plinius/status/2080782804435251373 ## Related - [[Claude 5]] - [[Claude Fable 5]] - [[Claude Sonnet 5]] - [[Claude Opus 4.7]] - [[Anthropic]] - [[Claude]] - [[Claude Code]] - [[Context Engineering]] - [[GPT-5.6]] - [[AI Frontier Model]] - [[Hacker News]]