⚡️ On May 28 Anthropic quietly rolled out Claude Opus 4.8 - no press conferences, no hyped-up pitch decks. It's not a revolution, it's a really smart upgrade that changes speed, cost and - most importantly - how the model behaves when it's not sure about the answer
What's new in Opus 4.8:
Context window up to 1M tokens, 128k max output. Available via Anthropic API, Amazon Bedrock, Google Vertex AI, Azure and GitHub Copilot (Pro+, Business, Enterprise). The price stayed the same: $5 per million input tokens and $25 per million output - same rates as 4.5, 4.6 and 4.7. A rare case in the AI industry where "it got better" doesn't come with a new bill
Concrete changes:
⚡️ Fast mode (research preview) - 2.5x faster than normal mode, $10/M input and $50/M output. If Opus 4.7 Fast mode felt like a private jet subscription - $30/$150 per million tokens, then 4.8 is at least business class
⚡️ Effort parameter - you pick the reasoning depth: low / high / extra / max. Want a quick answer - low, want the model to go full philosopher and rethink how the universe works - extra/max. Adaptive thinking decides on its own per-turn. No more setting the token budget by hand
⚡️ Honesty and self-correction - Anthropic claim Opus 4.8 lets its own bugs slide without comment 4x less often than 4.7. A "substantially lower" level of deceptive behavior in internal tests. The model asks more often when it's unsure and pushes back on a questionable plan by itself
⚡️ API upgrades - mid-conversation system messages (update instructions without resetting the cache), refusal stop details (the refusal category in the response), prompt caching minimum lowered to 1,024 tokens
What changes in practice:
💡 Agent sessions - fewer tokens lost on thinking, fewer skipped tool calls that matter, better recovery after compaction in long sessions
💡 Coding on big codebases - the model holds context better, loses the thread less and notices on its own when something's off with the logic
💡 Cost management - you choose how many resources to spend per task. Simple question - low effort, complex multi-step architecture - extra or max
Alongside the release Anthropic confirmed: Mythos-class models are coming "in the coming weeks" - Enterprise API first, then public. Mythos Preview scored 93.9% on SWE-Bench Verified and 97.6% on the USAMO 2026 olympiad math. Opus 4.8 isn't the final stop, it's the bridge to the next class
💼 With the release Anthropic closed Series H: $65B at a $965B valuation - higher than OpenAI ($852B). Annual revenue passed $47B by May 2026
What I run since then
Versions have changed since, and the brains of my 8 agents are now Sonnet 5 and Opus 5.5 for different roles. But the lesson from this release stayed: I don't take a model's honesty on faith, even when it's backed by metrics. In the content factory, names, numbers, dates and quotes are checked against the collected material before I even see the draft, and that has produced 84 posts, none without my approval. My personal agent Jes, running for 134 days, has everything risky locked behind rules: deleting data, spending money, messaging clients and publishing wait for my yes. A model can get four times more honest, and you still need checks around it
Advice for any new release: don't switch the whole system at once. Move one agent, run a dozen real tasks on it and compare with the previous model, and only then move on
🎯 Opus 4.8 is an upgrade that doesn't knock you out with the first announcement. Same price, Fast mode 3x cheaper than 4.7's, and for the first time Anthropic openly talk about honesty as a feature - with concrete metrics, not just "we're working on safety". If the claimed 4x on self-correction is real - this is the first model that can systematically say "not sure, let's check". Let's see how OpenAI respond