On Thursday, Anthropic released the newest version of its flagship model, Opus 4.8, which allows users to control Claude’s effort, tackle bigger coding tasks, and run fast mode more cheaply. Anthropic also says the model is more honest, less deceptive, and better at supporting user autonomy and best interests.
Benchmarks put Opus 4.8 ahead of its predecessor, as well as GPT-5.5 and Gemini 3.1 Pro, save agentic terminal coding, where OpenAI’s model remains the winner.
It was available on Thursday at the same price as Opus 4.7.
On May 28, X was atwitter about the supposed leak of Opus 4.8’s imminent arrival. One user posted a rumor that “Opus 4.8 has been found staged in the Claude Code model selector on the desktop app. It should be releasing today!”
And the tweets on the street were right, but how will Opus 4.8 be remembered in Opus history?
A new control means users can scale up or down how much elbow grease Claude puts into its tasks. When giving it its all, Claude will “think more frequently and more deeply to give a better response,” explains Anthropic in its announcement blog post. Conversely, lower-effort Claude will turn out faster responses and work through a user’s rate limit more slowly.
That could be good news for users feeling the effects of AI shrinkflation and worried about hitting rate limits faster than expected.
Now available in research preview, what Anthropic describes as a “dynamic workflows” feature should allow users to work on larger-scale problems with Claude Code. Specifically, Anthropic says users can now ask Claude to “plan the work and then run hundreds of parallel subagents in a single session.” From there, it will verify the outputs before returning them to the user.
What could that look like for an everyday user? Anthropic gives the example of codebase-scale migrations, saying Claude Code with Opus 4.8 can get it done “across hundreds of thousands of lines of code from kickoff to merge.”
More good news for Claude users’ wallets: Anthropic says Opus 4.8’s fast mode (i.e., when the model runs at 2.5x its normal speed) “is now three times cheaper than it was for previous models.”
That is, the model “reaches new highs on our measures of prosocial traits,” says Anthropic’s Alignment team in the announcement blog post. Specifically, the AI company reports that Opus 4.8 has improved support for user autonomy and for working in the user’s best interests.
In what looks like more good news, Anthropic says Opus 4.8’s rates of deception and cooperation with misuse are “substantially lower” than its predecessors, apparently catching up to Claude Mythos Preview, what the AI company once called “the best-aligned model we’ve trained,” reported by The New Stack.
Another big improvement for the model is its improved honesty. According to Anthropic, Opus 4.8 is “around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.” It claims that earlier testers have confirmed this statement and described Opus 4.8 as “more reliable and sharper in its judgment when it’s performing agentic tasks.”
Anthropic says Opus 4.8 levels up from its predecessor across all benchmarks. While launch-day benchmarks don’t always tell the same story as real-world use, the numbers do show promise.
While launch-day benchmarks don’t always tell the same story as real-world use, the numbers do show promise.
Most notable: Opus 4.8 shows a marked improvement over Opus 4.7 (64.3%) — not to mention GPT-5.5 (58.65) and Gemini 3.1 Pro (54.2%) — in agentic coding at 69.2%. Its agentic compute use score (83.4%) compared to GPT-5.5 (78.7%) and Gemini 3.1 Pro (76.2%) is nothing to sneeze at. But it loses out to GPT-5.5 in agentic terminal coding, down 3.6% compared to OpenAI’s model.
In May 2025, Anthropic launched Opus 4 at its first developer’s conference, Code with Claude, dubbing it “the world’s best coding model.” At the time, the AI company promised to set new standards for coding, advanced reasoning, and AI agents. The model brought significant advancements in coding and long-context reasoning, standing out for its ability to handle long-running tasks and maintain context in what Anthropic described at the time as “thousands of steps.”
Opus 4.1 followed soon after, in August 2025, with moderate improvements in the model’s performance on agentic tasks, coding, and reasoning. But it was only a small update; at the time, Anthropic teased of “substantially larger improvements to our model in the coming weeks.”
In November 2025, Opus 4.5 dropped, to much noise. Again, Anthropic touted it as “the best model in the world for coding, agents, and computer use.” And again — they teased us, noting Opus 4.5 was just “a preview of larger changes to how work gets done.” For its part, that preview brought improvements that enabled the model to better handle ambiguity and solve problems involving multi-system bugs. In many ways, Opus 4.5 allowed Anthropic to reclaim the coding crown after OpenAI’s GPT-5.1-Codex-Max and Google’s Gemini 3 model gained favor.
True to their hints, three months later, Anthropic gave us Opus 4.6, “a step change in using large language models (LLMs) for enterprise workflows, thanks to its ability to handle more complex tasks and deliver results better,” The New Stack reported. Opus 4.6 leveled up with better planning, coding, and debugging skills, became Anthropic’s first model to use adaptive thinking, and earned standout benchmark scores. Of particular note was its 1M-token context window. An Anthropic spokesperson told The New Stack, “it gets much closer to production-ready quality on the first try than what we’ve seen with any model – documents, spreadsheets, and presentations will need less back-and-forth on iterations.”
But all that glitters is not gold. On the heels of the Opus 4.6 launch, Anthropic caught flak for a pricing change: “While the models technically supported prompts approaching the 1-million-token limit, requests exceeding roughly 200,000 tokens were billed at higher ‘long-context’ pricing tiers, moving the entire request into a premium rate band,” The New Stack reported.
Opus 4.7 also faced some complications. After its drop in April 2026 — a direct upgrade to Opus 4.6 that brought better vision, better memory, and better instruction-following — The New Stack reported that “Claude Opus 4.7 users report self-contradicting responses and degraded performance, raising questions about AI model quality, safety tradeoffs, and shrinkflation.” Also awkward: Anthropic itself called Opus 4.7 “less broadly capable” than the then-much-talked-about Claude Mythos Preview. As The New Stack reported, Opus 4.7 appeared to be something of a training ground for new cyber safeguards for Mythos.
The rumors of a May-28 launch for Opus 4.8 proved right on the money, so it may be worth paying attention to the rest of the Internet gossip: the leak also suggested that Anthropic will soon announce Sonnet 4.8 and Mythos 1.
It would be big news, to say the least, from the AI company that has been frustrating users lately.
Earlier this month, it disappointed developers with Claude Code agent view, which, as Rob May, CEO and co-founder at Neurometric AI, told The New Stack: “It removes some friction, but it doesn’t change the underlying problem.” That same week, Anthropic also announced it will split billing for Agent SDK usage starting June 15, not exactly welcome news for users who were used to seeing programmatic usage and interactive usage pulled from the same subscription limits.
Perhaps Mythos 1 and Sonnet 4.8 will bring more wins.
YOUTUBE.COM/THENEWSTACK
Tech moves fast, don't miss an episode. Subscribe to our YouTube channel to stream all our podcasts, interviews, demos, and more.
Created with Sketch.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。