What actually changed in Anthropic's flagship model, what the new effort control does, and where Max effort earns its tokens. Released 28 May 2026, same price as 4.7.
"A modest but tangible improvement on its predecessor."
Anthropic's own framing. Translation for us: not a re-platform, not a reason to change how we work, but a genuine lift in coding accuracy, maths reasoning, honesty, and how independently it can run a long task. The bigger story is the new controls shipped alongside it.
Honest caveat. These are Anthropic's own tests. 4.8 leads most public benchmarks against GPT-5.5 and Gemini 3.1 Pro, but GPT-5.5 still leads on terminal and CLI coding, and the two are roughly tied on web browsing and graduate-level science. Opus 4.8 sits between 4.7 and the unreleased Mythos Preview on Anthropic's internal capability ladder.
There is now a control beside the model selector that lets me dial how hard Claude works on a single response. Higher effort means deeper thinking, more tokens, slower output, and better results. Lower effort means faster answers that are lighter on rate limits. This is the direct fix for 4.7's habit of overthinking simple tasks.
Fastest, cheapest on your limit. Quick lookups, reformatting, simple drafts, straightforward questions.
Everyday drafting, summaries, moderate analysis. The sensible middle for routine work.
Anthropic's default. Roughly matches 4.7's token spend on coding while performing better. Most substantive work sits here.
Slowest, most tokens, deepest reasoning. Hard multi-step problems, high-stakes accuracy, long autonomous runs.
A separate toggle that hands the decision back to the model, it judges how long to think per task. Useful when the workload is mixed and you do not want to keep changing the dial.
Worth confirming in your own selector. Labels differ slightly by surface. In Claude Code the high tier above the default is called "extra" or "xhigh", with "max" above that, and Anthropic has raised Claude Code rate limits to absorb the heavier token use. Early reporting also disagrees on the exact consumer default, so check what your model selector dropdown actually shows on claude.ai before you assume a setting.
In Claude Code, Claude plans the work then spawns hundreds of parallel subagents in one session. Anthropic's example is a codebase-scale migration across hundreds of thousands of lines, from kickoff to merge, with the existing test suite as the bar.
The Low to Max dial covered above. The single most useful day-to-day change for how we work in claude.ai and Cowork, because it puts cost and depth in our hands rather than the model's.
Fast mode runs Opus 4.8 at about 2.5x the speed and is reportedly three times cheaper than it was on previous models. Standard pricing is held flat at 4.7 levels.
The Messages API now accepts system entries inside the messages array. For our Make and MCP builds this means updating Claude's instructions mid-task without breaking the prompt cache or routing the change through a fake user turn. A quiet but genuinely useful change for automation.
Anthropic teased a more capable Mythos-class model for all customers in the coming weeks, pending safeguards. Worth tracking, but nothing to plan around yet. 4.8 is the working tool for now.
Sources. Anthropic launch announcement and system card (28 May 2026), as reported across Axios, Gizmodo, TechCrunch, VentureBeat, MacRumors, 9to5Mac, The Decoder, and Thurrott. Benchmark figures are Anthropic's own. Effort-level labels and the consumer default vary by surface and across early reporting, confirm in your model selector.