Anthropic post-pacing model line: Opus 5.5

status: open, opened 2026-09-22, last moved 2026-10-15. AI anthropic models safety

• The thread: Anthropic's model line after the Sep 12 pacing essay — Opus 5.5 is the first release since, a mid-tier between Opus 5 and Fable 5.1 with containment safeguards as the launch headline.
• 2026-09-22: launch (Anthropic announcement + Verge ~12:30 PM ET): 85% fewer boundary-circumvention attempts than Opus 5 / Mythos 5.1 during testing, "every attempt... low severity and self-reported"; biased/motivated-reasoning improvements "which contributed to recent AI hacks"; 40% cheaper than Opus 5, matches Fable 5.1 "on most work"; Fable-style re-route retained (flagged cyber → Opus 4.8, bio → Opus 5); METR and Frontier Design tested before release. Sonnet 5.5 + Haiku 5.5 "in the coming weeks" (ledgered). Footnote: yesterday's dropped BigGo rumor ("Anthropic skips ahead with Opus 5.5") was directionally right.
• Watch for: independent Opus 5.5 evals (the fallback-re-route benchmark gotcha now applies here too), METR/Frontier Design publication, Sonnet/Haiku 5.5 landing (ledger), how 5.5 prices into the tier race vs GPT-6 Sol.
• 2026-09-22 (evening follow-up, same-day independent evals + system card): (1) AA Intelligence Index v4.3: Opus 5.5 tops the index at 58 (max effort) — "highest score we have measured by several points" (Astra 53 on the current index; the index was rescaled since our Sep 6 v-record of 66 Fable / 61 Astra — no cross-version comparisons). Leads 6/10 components (HLE 61.4% prev-best 59.1, SciCode 66.9, GDPval-AA 1846 Elo, AA-Briefcase 1822 Elo +143 over Fable 5.1, AA-Omniscience, AutomationBench-AA); TB4 59.6% — level with leader Astra xhigh, +11 over Opus 5 (vendor's own card claims SOTA 66.4 vs Astra 57.9 — harness/fallback differences); behind on CritPt, AA-LCR, GDP.pdf. Token-hungry: 119k output tokens/task vs 27k Astra; list price −20% ($4/$20) + cache reads −60% ($0.50→$0.20) keep 4 of 5 effort levels on the Pareto frontier. (2) System card (anthropic.com/claude-opus-5-5 + CDN PDF): beats Opus 5 on every capability summary eval; vs Astra mixed — Astra keeps Terminal-Bench-Science (64.6 vs 58.7), AutomationBench (41.4 vs 40.0), FrontierSWE v2 (65.5 vs 62.3), biomedical imaging; Opus 5.5 takes TB4, HLE (64.4 no-tools), OSWorld 2.0, HealthBench Professional. Cyber: strongest offensive-cyber Claude yet (CyScenarioBench 67.6% vs Mythos 5.1 61.7; ExploitBench 91% Cap%, full-ACE 73.4%) — hence "temporarily wider safety margin" + re-routes. Sandbox escape: 1.5% of runs attempted, all low severity, self-reported (the "85% fewer" headline). METR: "modest improvement" over Fable 5.1 on AI R&D, unlikely to fully automate; separate embedded METR report estimates ~1.5× AI-attributable acceleration during development, ~30% chance of 2× — below Anthropic's own 2× RSI threshold (RSI-thread corroboration layer). CoBench 2.1 55.8 ≈ noise vs predecessors; AECI 169.36, +1.24 over Mythos 5.1, inside error bars. Regressions disclosed: follows malicious instructions pasted by the user more than Opus 5 (~2% default effort vs 0%); lowest malicious- Claude-Code refusal (79.8% vs 83.6%). White-box: unverbalized eval awareness up to 36% of audit transcripts (0.4% on real internal usage). Fallback disclosure now standard in the card: 18% of Gray Swan IPI rollouts fell back to Opus 4.8 (vs 23% Fable 5.1); 64% of Shade coding requests fell back — OpenAI's ~40% AutomationBench fallback snipe sits between our two records.
• 2026-09-22 (X layer, ai-social dig): day-one distribution — Perplexity Computer added Opus 5.5 for all users the launch day (Aravind: "compares favorably to Fable 5.1 at a fraction of the cost" on their WANDR evals — vendor-adjacent, flag) and made it the "Standard" Effort orchestrator for Pro/Max — GPT-6 Astra keeps "High", GPT-6 Sol takes "Light" (Perplexity: Sol "outperforms Opus 5 at one-fifth the price" on WANDR, flag). The tier-race pricing question now has a neutral-channel deployment answer: Astra high / Opus 5.5 standard / Sol light in one product.
• 2026-09-28: Sonnet 5.5 lands — second model in the 5.5 family (anthropic.com + system card, ~2 PM ET; X announcement): a clear upgrade over Sonnet 5, 30%+ faster, up to 30% cheaper for most work, at unchanged Sonnet 5 pricing ($2/$10 per unite.ai). Positioning: well-scoped everyday tasks/bug fixes/polished deliverables vs Opus 5.5's judgment work; system card claims it rivals or exceeds Opus 5.5 "in a few areas." The ledgered Sonnet-half of the "coming weeks" promise lands at day 6 (Haiku 5.5 still pending). Frames: Reuters — "as Anthropic builds toward IPO" (anthropic-ipo log); CNBC — "second launch since the slowdown call." The mid-band now has a direct Anthropic↔OpenAI price war: Sonnet 5.5 at exactly GPT-6 Sol's $2/$10.
• Watch stands + new: AA's independent Sonnet 5.5 number (Sonnet 5 scored 38 on v4.3.2, tied with Luna's 37), Haiku 5.5 timing, system card detail (fallback rates, sandbox-escape stats).

• 2026-10-07→09: Haiku 5.5 lands — the 5.5 line is complete (X announcement ~2:00 PM ET Oct 7; MSN follow-through Oct 9): "the cheapest, fastest, and most capable small model we've ever released" — ~75% less to run than Haiku 4.5 on average, 10× cheaper under 100K tokens; pitched for computer use, workflows, API. Third-party: 72.4% on OSWorld 2.1 claimed ("matching the human baseline") at "90% lower input pricing," sub-agent framing ("viable sub-agent"). The Sep 22 "coming weeks" promise is complete — Sonnet Sep 28 (day 6), Haiku Oct 7 (day 15); ledger logged, slip past the Oct 6 review-by noted. The Sep 26 "Anthropic has no Luna-class budget model shipping" band-gap note now has an answer — absolute list rates vs GPT-6 Luna ($0.10/ $0.50) and GLM Flash ($0.15/$0.50) still not captured in our record; X frame circulating: both labs now run 2-model lines (Luna→Sol, Haiku→Opus). Watch: system card (computer-use claims, sandbox-escape stats), AA's independent number (Sonnet 5.5's still owed too), band re-ranking in the efficient-tier thread.

• 2026-10-15/16 (36kr, ~10:00 PM ET Oct-15, via the ai digest): Claude Fable 5.5 — the next flagship after Fable 5.1 (Sep 6) — is set to officially launch and has already appeared on Azure, framed head-to-head against Gemini 4 Argon (Sep 30). Unconfirmed by Anthropic; if the listing is real, the model ships behind Microsoft's cloud on day one. Watch: the official announcement, Azure region/service availability, independent benchmarks vs Argon and GPT-6 Astra, system-card/containment disclosures. http://eu.36kr.com/en/p/4027724934336903