Claude Opus 4.8: the Opus-tier workhorse, upgraded

Claude Opus 4.8 is Anthropic's most capable Opus-tier model, released May 30, 2026 as a drop-in upgrade for Opus 4.7 — same $5 per million input tokens, same $25 per million output, same 1M-token context window. What changed is behavior, not the spec sheet: measurably stronger long-horizon agentic work (multi-step tool runs drift less) and noticeably warmer prose, addressing the most common complaint about 4.7's clinical register. Like the rest of the mid-2026 Claude line it uses adaptive thinking only — the model calibrates its own reasoning depth per request. Opus 4.8's position in the lineup shifted with the June launch of Claude Fable 5: Opus is no longer the top of the range but the price-performance anchor of Anthropic's premium tier — frontier coding and agent quality at half of Fable 5's $10/$50. The case for Opus 4.8: production coding, refactors, agent pipelines, and editorial writing. The case against: the very hardest reasoning tasks (Fable 5) and high-volume everyday work (Sonnet 5 at a third of the price).

Specs

Best at

• Production coding — the Opus line's refactor quality, now with fewer mid-run derailments in agent loops.
• Long-horizon agents — the headline 4.8 improvement: plans survive longer tool-call chains.
• Writing with voice — warmer prose than 4.7; the strongest editorial register below Fable 5.
• Whole-repo context — 1M tokens fits large codebases in a single call.

Where it loses

• Hardest reasoning — Claude Fable 5 sits a full tier above on research-grade problems.
• Cost at volume — Sonnet 5 delivers close-to-Opus capability at $3/$15 (intro $2/$10 through Aug 31, 2026).
• Fast shallow tasks — adaptive thinking can spend more reasoning than a lookup deserves; Haiku 4.5 is the speed pick.

Opus 4.8 vs Opus 4.7: what actually changed

Nothing on the invoice — pricing ($5/$25) and the 1M context window carry over exactly. The upgrade is in the weights: long-horizon agentic work is stronger, meaning multi-step tool runs hold their plan further, and prose is warmer, fixing the flat register 4.7 was known for. Opus 4.7 was deprecated from Council's active lineup in July 2026; existing conversations that used it still resolve, but new councils seat 4.8. There is no reason to prefer 4.7 — this is a strict, same-price improvement.

How Council AI runs it

Opus 4.8 is available on every paid Council AI plan (Plus, Pro, Ultra) and is one of the most-seated models in user councils — typically as the coding and writing anchor. One prompt fans out to Opus 4.8 and models from up to 8 other labs in parallel; each answers independently, then the moderator synthesizes a single response and scores how strongly the models agree. When Opus 4.8 and GPT-5.6 converge on the same patch, that consensus is a much stronger signal than either model's confidence alone.

Frequently asked questions