Claude Opus 4.8 review (Simon Willison) — source summary
Simon Willison's 28 May 2026 write-up of claude-opus-4-8. Delivered via Telegram,
ingested 2026-05-29. Text in raw/claude-opus-4-8-review.md.
What it says
- Anthropic frames the release, refreshingly, as "a modest but tangible improvement" over 4.7 — Willison's favorite part is the honest, low-hype framing.
- Headline improvement is honesty: Opus 4.8 is "around four times less likely than its predecessor" to make unsupported claims, and more likely to flag uncertainties about its own work. (Anthropic notes models often "jump to conclusions, confidently claiming progress despite thin evidence.")
- New capabilities he flags: mid-conversation system messages (
role: "system"after a user turn — steer instructions late in a long agentic loop without restating the prompt, preserving cache hits); lower prompt-cache minimum (1,024 tokens, down from 4,096 on 4.7). - Unchanged: $5/$25 per Mtok in/out (same as 4.5–4.7); Jan 2026 knowledge/training cutoff; 1M-token context; 128K max output. Fast mode now 2× base ($10/$50, down from $30/$150), research-preview only.
- Includes his signature "pelican riding a bicycle" SVG test across thinking levels.
Freshness check (2026-07-04)
Re-verified the volatile facts against Anthropic's current model catalog — the review's core
numbers still hold: claude-opus-4-8 remains the current Opus tier at $5/$25 per Mtok,
1M context, 128K max output, Jan 2026 cutoff. What has moved is the lineup around it:
the family expanded above and beside Opus 4.8 — Claude Fable 5 (now Anthropic's most-capable
widely-released model, priced above Opus at $10/$50) and its Project-Glasswing twin Claude Mythos 5,
plus Claude Sonnet 5. So the 28-May framing of Opus 4.8 as "the latest" is now one tier down from
Fable 5. Cross-spoke seam: Fable 5 / Mythos 5 are the very models in ../ai-governance-wiki's
export-control saga (the June ban / July re-release) — the substrate this
wiki tracks is the same object that spoke governs.
Flagged discrepancy (record-don't-overwrite). The review reports Opus 4.8's prompt-cache minimum "1,024 tokens, down from 4,096 on 4.7." Anthropic's current caching docs list Opus 4.8 at 4,096 tokens (1,024 is the Sonnet-4.5 tier; Fable 5 is 2,048). Either the launch-day figure changed back or was misreported — kept both readings; the review's number is a 2026-05-28 snapshot, not re-confirmed.
Why it's here
claude-opus-4-8 is the model substrate under the agent ecosystem this wiki tracks — claude-cowork, claude-managed-agents, and the Claude-driven implementations gbrain and llm-wiki-agent. Reflexively, it is also the model maintaining this very wiki, and its flagged improvement — flagging uncertainties rather than overclaiming — is precisely the discipline this wiki's synthesis and lint passes try to enforce.
Why this is T3 (re-tiered 2026-08-28)
Graded T4 on the masthead, and T4 was the wrong end. The load-bearing content here is Anthropic's own framing and Anthropic's own numbers — the "four times less likely" honesty figure is the vendor's — passed through a well-regarded commentator who did not measure anything. That is the ladder's middle case exactly: the vendor speaking through a writer, "usable for what's offered; flag promotional."
Not T2, because no test was run. The value here is the release framing and one observer's read of it.
Recovered by the recall probe (2026-08-30)
The probe run over this spoke's raw/ found two numbers in this source that no page here carried:
the context window is still 1,000,000 tokens and max output is 128,000. Unchanged from the
previous model, which is presumably why the ingest passed over them — and precisely why they were
worth keeping. "Still 1M" is a fact about a ceiling that did not move while the model changed, and
a corpus that records only what changed cannot answer whether anything stayed put. Cross-check:
../llm-providers-wiki holds the same figures for claude-opus-5 from Anthropic's own docs.