Anthropic released Claude Opus 4.8 on May 28, about 41 days after Opus 4.7 — a faster-than-usual cycle that likely reflects mixed reception to the previous version and pressure from OpenAI and Google. Anthropic describes it as “a modest but tangible improvement,” which is refreshingly honest for a product announcement.
Pricing stays the same: $5/M input and $25/M output tokens. Fast mode dropped significantly, now at $10/$50 per million tokens — about 3x cheaper than before, with 2.5x the speed.
<h2>What’s actually different</h2>
Honesty and reliability. The headline claim is that Opus 4.8 is 4x less likely to let code flaws pass unremarked. In practice, this means the model flags uncertainty, catches its own bugs, and abstains on questions it can’t answer confidently rather than hallucinating a response. On SWE-bench Pro, scores moved from 64.3% to 69.2%. Agentic terminal coding improved by over 8 points. Computer use scores 84% on Online-Mind2Web.
Several early adopters — Cursor, Cognition (Devin), Harvey, Databricks — report the reliability gains are real in production. Cursor noted it’s “more persistent on harder tasks.” Augment Code found it handles multi-hour executions and ticket-to-PR workflows with minimal intervention.
Effort controls. Users can now set thinking depth to low, medium, high, x-high, or max instead of manually managing token budgets. At low and medium, the model scopes its work to what was asked rather than going above and beyond — useful for controlling costs. For complex work, high and above is recommended. Anthropic says they put significant work into calibrating this, and the <code>/effort</code> command is available in Claude Code and Cowork.
Dynamic workflows. In research preview, Claude Code can now coordinate hundreds of parallel subagents for large-scale tasks like codebase migrations. This is the feature getting the most attention from developers. It burns tokens aggressively, so it’s not for casual use, but for genuinely large refactors it’s a meaningful capability shift.
Mid-task system prompt updates. The Messages API now supports updating system instructions mid-conversation without breaking the prompt cache. The minimum prompt cache size also dropped from 4,096 to 1,024 tokens.
Context window stays at 1M tokens with 128K max output. Knowledge cutoff is January 2026, same as 4.7.
<h2>Where it’s available</h2>
Opus 4.8 is live in <a href="http://Claude.ai" target="_blank" rel="noopener nofollow">Claude.ai</a>, Claude Code, AWS (via Amazon Bedrock and Claude Platform on AWS), GitHub Copilot (Pro+, Business, and Enterprise), Cursor, Windsurf, Devin CLI, Perplexity (Max subscribers), and Kilo Code.
<h2>Independent testing</h2>
One developer ran a 7-task benchmark covering frontend, 3D rendering, SVG generation, game building, math, and local fine-tuning workflows. Opus 4.8 scored 87.14% (61/70), up from Opus 4.7’s 55.71% (39/70), and ahead of GPT-5.5, Gemini 3.5 Flash, Deepseek, and Mimo. A combinatorics problem that every other tested model got wrong was a standout.
That said, a more skeptical hands-on review flagged that SWE-bench has reliability issues (up to 20% of passing runs potentially gaming the benchmark), that dynamic workflows burn tokens with mixed reliability, and that GPT-5.5 remains competitive. One developer noted the model “talks too much” and stuck with an older version.
The honest picture: it’s a real improvement over 4.7, especially for agentic coding and tasks where you want the model to flag problems rather than paper over them. Whether it’s a clear winner over GPT-5.5 depends on your specific workload.
<h2>What’s coming</h2>
Anthropic teased Mythos-class models — a full capability tier above Opus — which have been in restricted access since April due to cybersecurity concerns. Mythos can autonomously find zero-day vulnerabilities and has already identified over 10,000 high/critical software vulnerabilities through Project Glasswing with partners including Apple, Google, Microsoft, and AWS. General release is expected within weeks. Sonnet 4.8 may also be announced soon.
Separately, Anthropic announced a $65B Series H at a $965B valuation, with annualized revenue growing from roughly $1B at end of 2024 to an estimated $30B projected for 2026.

Collections

Anthropic released Claude Opus 4.8 on May 28, about 41 days after Opus 4.7. Pricing remains unchanged at $5/$25 per million tokens, while Fast mode dropped to $10/$50 — roughly 3x cheaper with 2.5x the speed. Key improvements include 4x better honesty/reliability (flagging bugs rather than hallucinating), SWE-bench Pro scores rising from 64.3% to 69.2%, configurable effort controls (low/medium/high/x-high/max), and dynamic multi-agent workflows in research preview that can coordinate hundreds of parallel subagents for large codebase tasks. The Messages API now supports mid-conversation system prompt updates without breaking cache, and minimum prompt cache size dropped from 4,096 to 1,024 tokens. Early adopters like Cursor, Devin, and Augment Code report real reliability gains in production. Independent benchmarks show meaningful improvement over 4.7, though GPT-5.5 remains competitive depending on workload. Anthropic also teased Mythos-class models — a tier above Opus — currently in restricted access due to cybersecurity concerns, with general release expected soon. The company announced a $65B Series H at a $965B valuation.

Claude Opus 4.8 released with longer autonomous work and improved self-assessment

Let’s hope this is a better upgrade than 4.7 was. Let’s go!!

Game on!! Let’s see what it’s worth