Claude Opus 4.8: Anthropic's New Flagship on a 42-Day Cadence
On 28 May 2026, Anthropic released Claude Opus 4.8 — just 42 days after Opus 4.7 (16 April 2026), the next point release in a tight cadence. Anthropic itself calls it a “modest but tangible improvement” and once again claims the top spot for the model. The headline for anyone using it in production: the token price stays unchanged, fast mode gets noticeably cheaper, and alignment sits close to the previously preview-only Mythos class, according to Anthropic. In parallel, Anthropic announced it will roll out that Mythos class “in the coming weeks” for all customers — including stronger coding and cyber capabilities.
What actually changed
- Released 28 May 2026, 42 days after Opus 4.7. Per Anthropic, the new number one — ahead of GPT-5.5 and Gemini 3.1 Pro in most tested categories.
- Price unchanged: 5 US dollars per million input tokens, 25 US dollars per million output tokens — identical to Opus 4.7.
- Fast mode: still 2.5x the speed, but per Anthropic now three times cheaper than for previous models — 10 US dollars input / 50 US dollars output per million tokens.
- Alignment substantially better than 4.7, per Anthropic: misbehavior at Mythos preview levels, and roughly four times less likely to let code flaws pass unremarked.
- SWE-Bench Pro: 69.2% (per Anthropic), ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro — but GPT-5.5 still leads on terminal coding.
What applied before
Opus 4.7 had been the Claude family’s flagship since 16 April 2026 — generally available via Claude.ai, the API, Bedrock, Vertex and GitHub Copilot. It brought a tangible improvement in software engineering over 4.6 plus better image understanding, kept the price of 5 US dollars input / 25 US dollars output per million tokens, but introduced a new tokenizer that can produce up to 35% more tokens for the same text. In effect, cost per task rose even though the per-token price stayed nominally the same.
Fast mode already existed in earlier versions and ran at 2.5x speed — but at a much higher token premium that made it uneconomical for many routine tasks. The Mythos class with its strongest alignment numbers was available only as a preview, not to the broad customer base.
What applies now
With Opus 4.8, Anthropic shifts three things in particular:
1. More performance at the same base price. Anthropic reports a SWE-Bench Pro score of 69.2% for Opus 4.8 — versus 64.3% for Opus 4.7 and 58.6% for GPT-5.5. On computer use, the model reaches 84% on Online-Mind2Web per the vendor, the highest score ever recorded on its in-house Legal Agent Benchmark, and is the only model — by Anthropic’s account — to solve every case in the Super-Agent benchmark end-to-end. That puts it ahead of GPT-5.5 and Gemini 3.1 Pro in most tested categories — with one clear exception: GPT-5.5 still leads on the terminal-coding benchmark. The token price stays at 5 US dollars input / 25 US dollars output per million tokens.
2. Cheaper fast mode. Fast mode still runs at 2.5x speed but, per Anthropic, now costs three times less than for earlier models. Concretely, the fast-mode price sits at 10 US dollars input / 50 US dollars output per million tokens — twice the standard price rather than the previously far steeper premium. For latency-critical agent workflows, that lowers the bar for using acceleration at all.
3. Alignment near Mythos level. Anthropic describes Opus 4.8’s misbehavior rates as substantially lower than 4.7’s and close to its best variant so far, Claude Mythos Preview. Practically relevant: per the vendor, the model lets code flaws pass unremarked roughly four times less often. The Mythos class itself — with even stronger coding and cyber capabilities — is slated to reach all customers “in the coming weeks,” once the safety measures are fully in place.
Alongside this, Anthropic launched a research preview of “Dynamic Workflows” for Claude Code, with hundreds of parallel subagents per session — a separate topic, noted here only as context.
Context
What stands out is less the individual model than the cadence: 42 days after 4.7, which itself came 41 days after 4.6. For agencies running Claude in production day to day, that means the flagship version number changes faster than an internal test-and-approval process can sensibly turn. Anyone evaluating every version individually can no longer keep up — it makes more sense to watch the few levers that actually matter in practice: price, speed, consumption.
On price, nothing moves, which is the good news — and also the catch. The base price has been stable since Opus 4.6; real consumption per task has not been: the 4.7 tokenizer alone produced up to 35% more tokens for the same text. So “price unchanged” does not mean “cost unchanged.” The genuinely new lever is the three-times-cheaper fast mode. Anyone who avoided acceleration because of the premium should run the numbers again — at 10 / 50 US dollars per million tokens, fast mode moves into the economical range for more tasks.
The Mythos outlook deserves the most sober reading. When Anthropic explicitly markets a model class around stronger cyber capabilities and ties the broad rollout to “safety measures,” that is a signal both ways: more capability to find defensive weaknesses — and more capability to exploit the same gaps offensively. For agencies with security-adjacent clients, this rollout is worth watching rather than dismissing as a marketing line. The alignment figures all come from Anthropic itself and are so far not independently verified; the same holds for the benchmark scores.
What you can do now
If you have avoided fast mode on cost: Run the numbers again. At 10 US dollars input / 50 US dollars output per million tokens and 2.5x speed, it can now come out cheaper than the old premium for latency-critical or iterative tasks — especially when faster responses raise throughput.
If you approve versions one by one: Switch to spot checks. Test Opus 4.8 on two or three representative tasks against your current setup instead of fully evaluating every point release. On a 42-day cadence, anything else is wasted time.
If your cost model is based on the token price: Measure real consumption, not just the per-token price. The unchanged base price says little while tokenizer and task complexity drive the output. A quick A/B comparison of token usage per standard task brings more clarity than the price list.
If you work in security or compliance: Set a reminder for the Mythos rollout. The announced cyber capabilities are both tool and risk — know the feature set before it goes live.
Background on the predecessor
What changed with Opus 4.7 — and why “price unchanged” still does not mean “cost unchanged”: → Opus 4.7: changes, impact, recommendations
Entdecke mehr
Opus 4.7 hasn't reached the quality level of 4.6 yet — our recommendations
What we see after several weeks on Opus 4.7 in practice — token appetite, overachieving, and concrete levers to push back.
BlogThree Days with Claude Fable 5 — and Why Opus Stays My Daily Driver
Fable 5 fixed a bug in one shot that Opus failed at twice. I'm still not switching. My honest impressions from the trenches after three days.
GlossarClaude Mythos
Claude Mythos (April 2026) is Anthropic's most powerful model to date — a new tier above the Opus line, distributed primarily through the Glasswing program to cybersecurity defenders.