Claude Opus 4.8 lands with dynamic sub-agent workflows and a welcome dose of honesty
Anthropic released Claude Opus 4.8 on May 28, the latest upgrade to its flagship model family. On paper it’s a point release — the API name is simply claude-opus-4-8 — but the launch bundles three structural changes that signal where the company is taking its product strategy.
What’s new
Opus 4.8 keeps the same pricing as Opus 4.7 ($5 per million input tokens, $25 per million output), which in a market where frontier model costs are drifting upward is a statement of intent. Fast mode — running at 2.5× speed — is now three times cheaper than it was on the previous generation, at $10/$50 per million tokens (Anthropic announcement).
The headline feature is dynamic workflows in Claude Code, Anthropic’s agentic coding environment available on Enterprise, Team, and Max plans. Instead of working linearly, Claude can now plan a task, spawn hundreds of parallel sub-agents, have them work simultaneously across a codebase, verify their outputs against the test suite, and return a consolidated result. Anthropic says this lets Claude “carry out codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge” (Anthropic, May 28). Early tester CursorBench reported that Opus 4.8 used fewer tool steps than its predecessor to achieve the same output (AI News, May 29).
The honesty improvement
Perhaps the most interesting change isn’t a benchmark score but a behavioural one. Anthropic’s alignment evaluation found Opus 4.8 is “around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.” The model is more prone to flag uncertainty and push back when it doesn’t have enough evidence for a claim. The company’s alignment team concluded the model “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest,” with rates of deception and cooperation-with-misuse comparable to Claude Mythos Preview, Anthropic’s strongest-aligned model (Anthropic system card).
Benchmarks
On SWE-bench Verified, Opus 4.8 scores 88.6%, leading the published four-model comparison against GPT-5.5 and Gemini 3.1 Pro across every SWE-bench variant (llm-stats.com; Digital Applied). It posts 74.6% on Terminal-Bench 2.1 and 1,890 Elo on GDPval-AA, a measure of general practical knowledge work. On Anthropic’s internal Super-Agent benchmark, it is the only model to complete every case end-to-end (Anthropic).
Effort control
A new “effort” dial on claude.ai and in Claude Code lets users trade token burn for quality. The default “high” spends about the same tokens as Opus 4.7’s baseline but performs better. “Extra” and “max” modes spend more tokens for harder tasks or async workflows. Anthropic has raised Claude Code rate limits to accommodate the higher ceiling (Anthropic).
What’s next
Anthropic used the Opus 4.8 launch to preview Project Glasswing, a group of organisations using Claude Mythos Preview for cybersecurity scanning. The company said it expects to bring “Mythos-class models” to all customers “in the coming weeks” (Reuters, May 28). The timing is notable: Anthropic also recently raised $65 billion at a valuation nearing $1 trillion ahead of a reported IPO (TechCrunch, May 28).
The big picture
Opus 4.8 is not a leap — it’s a refinement. The pricing freeze suggests Anthropic is fighting to hold share against OpenAI and Google on cost, while the dynamic workflows feature is a bet that enterprise customers don’t just want a smarter chatbot; they want infrastructure for autonomous code work at scale. The honesty improvement, meanwhile, addresses a real pain point: models that confidently deliver wrong output waste more developer time than models that stop to say “I’m not sure.” Whether those behavioural gains persist under adversarial pressure will matter more than any benchmark chart.
Sources: Anthropic official announcement (May 28, 2026); AI News (May 29, 2026); Reuters (May 28, 2026); llm-stats.com; Digital Applied; TechCrunch (May 28, 2026)