What shipped on May 28<\/h2>\n
Per Anthropic’s announcement<\/a>, Claude Opus 4.8 launched on May 28, 2026 across the API, claude.ai, and Claude Code, under the model ID Three things headlined the release:<\/p>\n There is also a quieter feature with daily consequences: adjustable effort levels, now surfaced in claude.ai as well as the API. You can dial the model down for routine tasks and save tokens, or up for the problems that deserve the full treatment. The default is high.<\/p>\n Six weeks is a short gap between flagships, and the API surface did not change at all, which means upgrading is a model-string swap. The improvements are behavioral, and two of them matter more than the benchmark table.<\/p>\n First, the honesty shift.<\/strong> According to Anthropic’s system card, Opus 4.8 is roughly four times less likely than its predecessor to let flaws in its own code pass unremarked. Independent reviews found a related pattern on factual benchmarks: the model posted the lowest incorrect-answer rate of six models tested, largely because it abstains when it is not sure instead of guessing, per Build Fast With AI’s review<\/a>. Early testers quoted in the announcement describe a model that “catches its own mistakes” and “pushes back when a plan isn’t sound.” For anyone who has shipped an AI-written bug because the model sounded confident, this is the feature.<\/p>\n Second, the benchmark jumps are real but uneven.<\/strong> The same review reports 88.6 percent on SWE-Bench Verified and 69.2 percent on SWE-Bench Pro, a 4.9-point gain on the harder test in six weeks, plus a leap on competition math (USAMO 2026, from 69.3 to 96.7 percent). Reviews place it ahead of GPT-5.5 and Gemini 3.1 Pro on most measures, with agentic terminal coding remaining contested territory. As always, benchmark leads of a few points matter less than whether the model finishes your tasks, and the early tester reports about end-to-end completion are the more relevant signal.<\/p>\n Third, dynamic workflows change the shape of what one session can do.<\/strong> Claude writes a JavaScript orchestration script for the task you describe, then a runtime executes it in the background, fanning work out across parallel subagents. Anthropic’s example is a codebase migration across hundreds of thousands of lines, validated against the existing test suite. It is a research preview, gated to paid team tiers, and it will be rough at the edges. It is also the clearest look yet at where agent tooling is headed: you describe the outcome, and the model manages the workforce.<\/p>\n Most readers of this site are not migrating half-million-line codebases. Here is where this release lands for a small operation.<\/p>\n The practical ceiling on agent work has never been intelligence. It is drift: the agent goes wrong at step 14 and burns an hour of tokens confidently building on the mistake. A model that is measurably better at flagging its own errors and stopping to say “this is not working” changes how long you can leave it alone. Multi-step jobs like “audit these 40 product pages and fix the broken schema markup” or “reconcile this export against last month’s” become things you start before lunch and review after, rather than supervise throughout.<\/p>\n Before adjustable effort, cost control meant juggling different models with different behaviors. Now the same model does the cheap version and the expensive version of thinking. Routine summarization runs at low effort; contract review runs at high. One model to learn, one bill to read.<\/p>\n At $30\/$150, fast mode was a curiosity. At $10\/$50 it is a tool for anything interactive, live drafting sessions, quick iteration on code, working meetings where waiting kills the flow. It is still double the standard price, so reserve it for work where speed is the point.<\/p>\n Research-heavy drafting is exactly the workload where the honesty improvement pays: a model that abstains rather than inventing a statistic is worth real money to anyone publishing under their own name. It is the same reason our Blogging System<\/a> pairs an AI drafting engine with a human editor before anything goes live, at $25 a month or $197 a year. This blog is drafted with Empower Network’s AI content engine and edited by a human before publishing.<\/p>\n It is not Anthropic’s best model.<\/strong> Claude Fable 5, released in June 2026, sits above it at double the price and scores more than 10 percent higher on some benchmarks across the hardest work. Opus 4.8 is the everyday tier, and Anthropic prices it accordingly. If your problem genuinely needs the ceiling, this is not the ceiling.<\/p>\n Dynamic workflows are a preview, not a product.<\/strong> Enterprise, Team, and Max only, and research previews change or disappear. Build habits on it, not businesses.<\/p>\n $25 per million output tokens is still flagship money.<\/strong> High-volume, low-difficulty workloads, like classification, extraction, and templated drafts, will quietly bleed budget here. Cheaper tiers exist for a reason.<\/p>\n Abstention has a flip side.<\/strong> A model tuned to say “I’m not sure” will sometimes hedge where you wanted an answer. For decisive, low-stakes generation, the caution can read as friction. Prompting for a definite recommendation usually resolves it, but it is a behavior change from 4.7 you will notice.<\/p>\n Pass on Opus 4.8 if your AI use is chat, drafting, and quick questions; the default models on free and consumer plans cover that, and flagship pricing buys you nothing. Pass if your workload is high-volume and simple, where a budget model finishes the same tasks at a fifth of the cost. And pass, at least on upgrading day-one habits, if you are already deep into a working Opus 4.7 setup and none of your tasks fail from the model catching errors too rarely; the swap is trivial, but re-tuning prompts for a slightly different temperament still takes an afternoon you may not need to spend.<\/p>\n Think of the current market as three tiers. At the top, ceiling models like Claude Fable 5<\/a> for the hardest, highest-stakes work. At the bottom, volume models like Claude Sonnet 5<\/a> for cheap, parallel, routine tasks. Opus 4.8 owns the middle: the daily workhorse you point at substantial agentic jobs, long coding sessions, document-heavy analysis, anything where an unflagged error costs more than the token bill. A sensible small-business stack uses all three deliberately, routing by task rather than by loyalty, and our agent setup guide<\/a> shows how to wire that routing without an engineering team.<\/p>\n The 41-day release gap is its own signal. Anthropic is now shipping flagship updates at the pace the rest of the market ships point releases, which means the model you standardize on today will be superseded within a quarter. That is an argument for building workflows that survive model swaps, and for judging each release the way this one deserves to be judged: not by its benchmark table, but by whether the failures that used to cost you time now announce themselves. On that measure, Opus 4.8 earns the upgrade.<\/p>\n This post was drafted with Empower Network’s AI content engine and edited by a human before publishing.<\/em><\/p>\n Claude Opus 4.8 shipped May 28, 2026 at unchanged $5\/$25 pricing, with 4x fewer unflagged code errors and parallel subagent workflows. What it changes.<\/p>\n","protected":false},"author":0,"featured_media":13,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[2,6,9],"class_list":["post-12","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","tag-ai-agents","tag-anthropic","tag-claude-opus-4-8"],"_links":{"self":[{"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/posts\/12","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/comments?post=12"}],"version-history":[{"count":2,"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/posts\/12\/revisions"}],"predecessor-version":[{"id":63,"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/posts\/12\/revisions\/63"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/media\/13"}],"wp:attachment":[{"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/media?parent=12"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/categories?post=12"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/empowernetwork.com\/ai\/wp-json\/wp\/v2\/tags?post=12"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}claude-opus-4-8<\/code>. Pricing held at $5 per million input tokens and $25 per million output, unchanged from Opus 4.7. The model keeps the 1M-token context window and 128K output ceiling of its predecessor.<\/p>\n\n
![\"Claude](\"https:\/\/empowernetwork.com\/wp-content\/uploads\/sites\/904\/2026\/07\/claude-opus-4-8-agentic-workhorse-infographic-scaled.png\")
<\/figure>\nWhat actually changed from Opus 4.7<\/h2>\n
What Claude Opus 4.8 means for an operator’s actual work<\/h2>\n
Longer runs without babysitting<\/h3>\n
An effort dial instead of a model switch<\/h3>\n
Fast mode at a defensible price<\/h3>\n
Content and research pipelines<\/h3>\n
The honest limits<\/h2>\n
Who should ignore this release<\/h2>\n
Where Opus 4.8 fits in the agent stack<\/h2>\n
Sources<\/h3>\n
\n