Anthropic released Claude Opus 4.8 on May 28, 2026 calling it "a modest but tangible improvement on its predecessor." For context, the new Opus came out just 41 days after the previous generation, which some users weren’t particularly happy with. But it also adds some very powerful new features — and we’ll explain them all in detail in this article.
Chat with Opus 4.8 on Overchat AI by following this link.
Claude Opus 4.8 is Anthropic's flagship model and a direct upgrade to Opus 4.7.
It came out quickly after Opus 4.7, which had been criticized for giving way to many verbose comments in generated code, and a tendency to over-think simple tasks.
Cognition, the team behind the Devin coding agent, explicitly said in Anthropic's launch post that Opus 4.8 "fixes the comment-verbosity and tool-calling issues we saw with Opus 4.7."
The model also shows improvements that are difficult to capture with traditional benchmarks, particularly in alignment.
What is alignment? It’s the degree to which an AI behaves in accordance with human intentions, follows instructions reliably, and responds safely in complex situations.
According to Anthropic's alignment researchers, Opus 4.8 performs nearly as well in these areas as Claude Mythos Preview, the company's most advanced internal model.
As a reminder, Mythos Preview is still available only to a small group of partners because Anthropic believes its capabilities could pose cybersecurity risks if they were to open access to the wider public.
If you're already using Opus 4.7, here's what to expect from the new model, both in terms of its visible features and the improvements under the hood.
Let’s talk about the new features added in this release in more detail:
Inside Claude Code, Opus 4.8 can now delegate large tasks to hundreds of subagents running in parallel, similar to the Agent Swarm capability introduced in Kimi K2.5/2.6.
According to Anthropic, this enables workflows that were previously difficult to automate, including codebase-wide migrations. One example they highlight is updating a codebase containing several hundred thousand lines of code from initial planning through to a merge-ready pull request, using the existing test suite as the primary measure of success.
This feature is available on Claude Code Enterprise, Team, and Max plans. Keep in mind that running hundreds of agents simultaneously can consume tokens and usage limits extremely quickly, especially on large or complex projects.
This is a new tier that:
This is useful for high-throughput workloads where speed matters more than per-token cost.
A small but practical change to the Messages API.
Unchanged from Opus 4.7. A few things to note here:
Opus 4.8 is around four times less likely than Opus 4.7 to ship a bug silently. Instead, if the model makes a mistake or generates code with a bug, Claude will let you know that it’s not sure if the code works.
Opus 4.7 in those instances were most likely to deliver flawed code but report it as working.
Second, Anthropic's alignment team gives Opus 4.8 a misalignment score of 1.9, down from 2.5 for Opus 4.7. That's also close to Claude Mythos Preview, which scores around 1.7 (lower scores are better.)
They assign those numbers based on roughly 2,600 simulated investigation sessions per model, where they measure deception, assisting with misuse, and ignoring user intent. The full evaluation spans 244 pages in the Opus 4.8 System Card. The short version: this is the most reliable Claude Anthropic has made available through its public API.
Opus 4.8 delivers its largest improvements in coding, mathematical reasoning, and agentic computer use. Here's the full benchmark table showing how it performs across the major evaluations:
Opus 4.7 was released six weeks earlier and was Anthropic's flagship until Opus 4.8 shipped. It introduced the multi-tier effort system, a 1M context window, and the agentic improvements that Opus 4.8 now builds on. The criticism that landed quickly: verbose code comments, clumsy tool calls, and a tendency to spend too many tokens on tasks that should have been quick.
The +4.9 jump on SWE-Bench Pro is the practical coding signal — that's the harder, less-saturated coding benchmark, and a 5-point gap means Opus 4.8 resolves real-world coding tasks Opus 4.7 couldn't. The +27 jump on USAMO 2026, the US Mathematical Olympiad, is the largest single-cycle math gain in the Opus line and signals a qualitative shift in mathematical reasoning depth. GPQA is flat — both models sit above the 93% saturation line where the remaining questions are at the edge of human expert capability.
GPT-5.5 is OpenAI's current top-tier model and sits at the top of the Artificial Analysis Intelligence Index at 60.2. It's the strongest model on terminal-style coding (where commands run live in a shell) and one of the cheapest flagships on output tokens.
The pattern is split. GPT-5.5 still wins Terminal-Bench by a clear 3.6 points, which matters for command-line-heavy work — running shells, package managers, system tools. Opus 4.8 wins by larger margins on SWE-Bench Pro (a 10-point lead, meaning it resolves real GitHub issues at a substantially higher rate), Humanity's Last Exam (the hardest reasoning test currently in use), agentic computer use, and GDPval — Artificial Analysis's measure of real economic-value tasks where a 121-Elo lead translates to roughly two-thirds head-to-head win rate. GPT-5.5 is 40% cheaper on output. For most knowledge-work and structured engineering, Opus 4.8 leads. For terminal-heavy CLI workflows or cost-sensitive output, GPT-5.5 still has a case.
Gemini 3.5 Flash is Google's mid-tier model, released May 19, 2026. It's positioned at "frontier intelligence at Flash latency" — roughly 4x faster than Opus and a third the price, but it trades raw capability for speed. Its native multimodality (text, image, audio, video, PDF) is a class apart from anything in the Opus line.
These models target genuinely different use cases. Opus 4.8 leads on every reasoning, coding, and knowledge benchmark in this set — and by margins large enough that for hard work it's the only sensible pick. But Gemini 3.5 Flash holds a small edge on MCP tool use, runs roughly 4x faster, and costs about a third as much. For high-volume work where throughput matters more than peak quality — drafting, summarizing, simple tool-use chains — Gemini Flash often delivers more value per dollar. For multimodal work, it's the better choice regardless.
The standard pricing is unchanged from Opus 4.7. Here’s the price per 1M tokens:
If you want to skip the API pricing entirely, you can chat with Claude Opus 4.8 on Overchat AI as part of a single subscription that also includes GPT-5.5, Gemini 3.5 Flash, Kimi K2.6, Qwen 3.7 Max, and more.
GPT-5.5 is still better for terminal work. On Terminal-Bench, GPT-5.5 scores 78.2 compared to Opus 4.8's 74.6. If you spend most of your time in the command line, GPT-5.5 still has the edge.
GPQA Diamond is slightly worse than before. Opus 4.8 scores 93.6, down from 94.2 on Opus 4.7. The difference is small, but performance did go down.
Output costs more than GPT-5.5. Opus 4.8 costs $25 per million output tokens, while GPT-5.5 costs $15. If your applications generate a lot of text, the higher price can add up quickly.
Anthropic is shipping new models very quickly. Opus 4.8 arrived just 41 days after Opus 4.7. If your team recently upgraded to 4.7, you may not be eager to switch again so soon, especially since prompts may need adjustments between versions.
Choose Opus 4.8 if:
Choose another model if:
Claude Opus 4.8 was released on May 28, 2026, just 41 days after Opus 4.7. It is available now through Overchat AI, Claude.ai, Claude Code, the Claude API, Amazon Bedrock, and Vertex AI.
Pricing is unchanged from Opus 4.7:
Anthropic also introduced Fast Mode, which runs at about 2.5× the normal speed. It costs $10 per million input tokens and $50 per million output tokens. Fast Mode is currently in preview.
For most real-world tasks, yes.
It scores higher on coding, math, and agent-based tasks. Some benchmarks stayed flat or dropped slightly, but those tests are already near the limit for today's top models.
The bigger improvement is reliability. Opus 4.8 produces far fewer unflagged coding errors and performs better on Anthropic's alignment tests.
Opus 4.8 performs better on coding, reasoning, and agent-based tasks. It also scores higher on benchmarks such as SWE-Bench Pro and Humanity's Last Exam.
GPT-5.5 still performs better on terminal-focused tasks and command-line workflows. It is also cheaper, costing about 40% less for output tokens.
In short:
Dynamic Workflows is a new feature in Claude Code. It lets Opus 4.8 break large tasks into smaller pieces and run them across hundreds of subagents at the same time.
Anthropic says this can be used for tasks such as large codebase migrations, where the model updates a project, runs tests, and prepares changes for review.
The feature is currently available on Claude Code Enterprise, Team, and Max plans.
Yes. Opus 4.8 is also strong at math, research, long-context tasks, and web-based agent workflows. Its benchmark results in these areas are among the best currently available.
With Claude Opus 4.8, Anthropic has fixed many of the verbosity and tool-use concerns that users raised with Opus 4.7. The model also produces code with far fewer unflagged bugs and is generally more reliable. At the same time, Anthropic's alignment team says Opus 4.8 performs close to Claude Mythos Preview, the company's most advanced internal model. That's one of the strongest alignment results the company has reported for a public model. To test Opus 4.8 on your own workflows today, head to Overchat AI and start chatting with Claude Opus 4.8.
