The 30-second answer: Claude Opus 4.8 is Anthropic’s new Opus model, released on May 28, 2026. The key changes versus 4.7 are a roughly one-quarter reduction in the rate of letting code defects through, an honesty improvement that avoids unsupported claims, and the addition of a fast mode that’s about 2.5x faster and 3x cheaper. The standard pricing ($5/$25) stays the same.
I keep Claude open all day and run my solo-business operations through it. So whenever the model steps up to a new version, the first thing I look at isn’t the flashy marketing copy — it’s “what changed versus the previous version” (the changelog). It was no different when Anthropic released Claude Opus 4.8 on May 28, 2026. This post is a release summary of those changes, based on the official announcement. (How a 1M context changes things for a solo developer in practice is covered in a separate post, so here I’ll only touch on it briefly.)
What launched, when, and where
Claude Opus 4.8 went public on Thursday, May 28, 2026. Its API model ID is claude-opus-4-8, and Anthropic positions it as “the most capable model generally available today.” Complex reasoning and long-horizon agentic coding are the emphasis points.
You can use it on these channels: the Claude API, Amazon Bedrock, Vertex AI, Microsoft Foundry, plus claude.ai and Claude Code. Note that Microsoft Foundry caps the context at 200k. The 1M (one-million) token context is available by default at standard pricing with no separate beta header; the maximum output is 128k (300k on Batch), and the knowledge cutoff is January 2026.
As of May 29, 2026, this model is the latest version in Anthropic’s Opus line.
The 5 key changes from 4.7 to 4.8
This is the heart of the post. Based on Anthropic’s announcement, here are the five things that changed versus 4.7.
1) Code quality — Opus 4.8 is about one-quarter as likely (roughly a 4x reduction) to let defects in its own code slip through, compared with the previous generation. This is the change that hits developers most directly, because it points toward less burden on the human reviewing the model’s code.
2) Honesty improvement — it’s more inclined to flag uncertain parts and avoid unsupported claims. Its rate of misaligned behavior is also lower than 4.7’s. That’s a meaningful change for work where the accuracy of information matters.
3) Longer autonomous work — it works independently for longer than the previous generation. Its ability to carry a long coding task forward without mid-way intervention is stronger.
4) Fast mode — a mode that’s about 2.5x faster and 3x cheaper was added. The fast mode price is $10 input / $50 output per million tokens — 3x lower than the prior cost structure ($30/$150). The standard price is $5 input / $25 output, the same as 4.5 through 4.7. For operators who fret over token-cost optimization, it amounts to a new option.
5) New features — Dynamic Workflows (research preview) was added: a feature for managing hundreds of subagents in parallel within a single session. There are also mid-conversation system messages now, letting you insert system instructions partway through a conversation.
Generation comparison table
Here are the recent Opus generations as of May 29, laid out in a table. Standard prices are per million tokens, input/output.
| Model | Context | Standard price (input/output) | Released |
|---|---|---|---|
| Opus 4.8 | 1M | $5 / $25 | 2026-05-28 |
| Opus 4.7 | 1M | $5 / $25 | — |
| Opus 4.6 | 1M | $5 / $25 | — |
| Opus 4.5 | 200k | $5 / $25 | — |
| Opus 4.1 | 200k | $15 / $75 (old price) | — |
The table shows that standard pricing has held at $5/$25 from 4.5 through 4.8, and the context expanded to 1M starting with the 4.6 generation. Compared with 4.1’s old pricing ($15/$75), the cost structure has come down sharply even within the same Opus line.
Who should switch now
Anthropic set 4.8’s effort default to high and recommends xhigh for coding work. It announced that migration is a matter of swapping the model ID with no breaking changes (the detailed procedure is covered in a separate post).
From my operating standpoint, the drop in code-defect pass-through and the honesty improvement are changes that cut “review time,” so a solo operator feels them keenly. Fast mode opens room for token-cost optimization in workflows with frequent bulk calls. Conversely, if you only do light one-off tasks, there’s little reason to rush.
One thing to be clear about, though: changes like these add new options across Claude Opus, Anthropic, and the developer ecosystem at large — they don’t produce the results for you. Choosing what to use through tool comparison, and planning and deciding what to delegate to the model, is still your job. The model makes things possible; what you fill it with is up to you.
What’s next in the series
This post is part 1 of a five-part Claude Opus 4.8 series (the release summary). Coming up: (2) coding benchmarks, (3) pricing and efficiency, (4) competitor comparison, and (5) migration.
Specs and pricing reflect the official announcement and may change, so check the official pages before adopting it.
Sources: Anthropic — Introducing Claude Opus 4.8, Claude Docs — Models overview, Claude Platform — Pricing