English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Opus 5.5 Arrives with Price Cuts, GPT-6 Sol Follows Within an Hour: The AI Coding Cost Curve Gets Rewritten

Forum topic · 小凯 · 2026-09-25

Summary

On September 22, 2026, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens—20% cheaper than Opus 5—with cache reads dropping 60% to $0.20. Anthropic claims a 40% net cost reduction on typical coding workloads, while output speed improved over 30%. Adaptive thinking is now a forced default that developers cannot disable via API. Independent benchmarks from Artificial Analysis complicate the picture: at max effort, Opus 5.5 generates roughly 60% more tokens, making per-task cost roughly flat versus Opus 5. One hour after launch, OpenAI quietly added GPT-6 Sol and Luna to Codex CLI via Amazon Bedrock, signaling next-generation flagships in production. Additional changes include a 20% higher Claude Code five-hour usage cap, banked rate-limit resets, and thinking blocks now bound to sessions to block distillation. Anthropic also disclosed the model frequently recognized it was being evaluated, and high-risk cyber/bio requests fall back to Opus 4.8 pending verification programs.

Opus 5.5 Arrives with Price Cuts, GPT-6 Sol Follows Within an Hour: The AI Coding Cost Curve Gets Rewritten

> [AI coding · flagship new generation] On September 22, 2026, Anthropic released Claude Opus 5.5 at $4/million input tokens and $20/million output tokens—each 20% below the previous Opus 5. Cache reads dropped from $0.50 to $0.20, a 60% saving on that line alone. One hour later, OpenAI quietly wired GPT-6 Sol and GPT-6 Luna into Codex CLI. This wasn't just another leaderboard refresh: the economics of coding agents were officially put back on the table.

One-sentence conclusion

By Anthropic's own reporting, Opus 5.5 matches its Fable 5.1 on most tasks, but charges only 40% of Fable 5.1's per-token price and 8% on cache reads; net cost drops about 40% and output speed is up 30%+. It is also the first Anthropic model to make "adaptive thinking" a forced default—developers can no longer switch off deep reasoning via the API.

Why this moment is different

For two years, each flagship release was judged almost entirely on benchmarks: SWE-bench, Terminal-Bench, Aider, AIME, nudged up by fractions of a point. This September the tone shifted—Anthropic's front page was filled with "40% cost reduction," "long agent runs no longer collapse," "no longer forgets the original plan after the first hour." The bet is straightforward: dev teams won't pay more for a slightly smarter model, but they will extend loop lengths 3x–5x when per-task cost falls.

OpenAI's response changed too. Same day, the Codex CLI release notes added GPT-6 Sol and Luna in a single line—but that line means Codex treats the new models as first-class citizens, not options buried in a dropdown.

The pricing, itemized

| Model | Input \(/M | Output\)/M | Cache read $/M | |---|---|---|---| | Fable 5.1 | 10 | 50 | 0.25 | | Opus 5 (previous) | 5 | 25 | 0.50 | | Opus 5.5 | 4 | 20 | 0.20 | | Opus 5.5 Fast mode | 8 | 40 | — | | Sonnet 5 | 2 | 10 | 0.20 | | Haiku 4.5 | 1 | 5 | 0.10 |

Cache reads matter enormously for coding agents: the loop repeatedly re-feeds repos, specs, and prior tool outputs into context—the same text read hundreds of times. Savings there are real money.

Performance figures and the third-party picture

| Dimension | Opus 5.5 self-reported | Comparison | |---|---|---| | Terminal-Bench 4.0 | 66.4% | above GPT-6 Astra's 57.9% | | GDPval-AA v2.1 (Elo) | 1846 | above Fable 5.1, Opus 5 | | Artificial Analysis Index (max effort) | 58 | +7 points vs Opus 5 | | Output tokens / AA task | ~119k | ~60% more than Opus 5 |

Third-party accounting cools the optimism: Artificial Analysis found that at max effort, the ~60% extra output tokens leave weighted per-task cost roughly flat versus Opus 5. Anthropic's 40% cost drop is a "typical coding workload" figure, not a max-effort one.

Adaptive thinking, forced on

Opus 5.5 makes adaptive thinking mandatory—developers can only modulate depth via an effort parameter. Anthropic's internal tests: an HAProxy C-to-Rust translation finished in 9.5 hours (Opus 5: 12 hours, 51% cost savings); a 200k-line codebase audit ran under 3 hours (Opus 5: 20 hours and 2.5x tokens).

Anthropic also admitted Opus 5.5 frequently recognized it was being evaluated during assessment—an open question for behavioral transfer to real deployment. Safety documentation cites Frontier Design and METR evaluations; general cybersecurity requests automatically fall back to Opus 4.8, with higher capability tiers gated behind the Cyber Verification Program or Life Sciences Verification Program.

GPT-6 Sol and Luna follow close behind

Codex CLI's release notes on the night of September 22:

> Added support for new GPT-6 Sol and Luna models via Amazon Bedrock, including migration prompts for older models.

The same release included AWS Bedrock access and old-model migration prompts—meaning production-ready invocation. Anthropic announces; one hour later, OpenAI ships two variants of its next-gen flagship into a coding tool. Sol is the flagship workhorse; Luna is likely the lightweight variant. This "opponent cuts prices, so do I" sync last happened at GPT-5 vs Claude Sonnet 4; now it's next-gen flagship level.

The wider view

Claude Code's five-hour usage cap rises 20% on Opus 5.5, with banked rate-limit resets introduced for the first time—save today's quota for tomorrow's heavy run. Availability spans Pro, Max, Team, and Enterprise plans; AWS Bedrock, Google Cloud, and Microsoft Azure all launched day one. Sonnet 5.5 and Haiku 5.5 are slated for the coming weeks.

One under-emphasized but consequential detail: thinking blocks are now bound to a single session, aimed at blocking model distillation. Competitors hoping to clone Claude's reasoning style lose access—and developers with multi-session workflows must rework them.

Boundaries worth noting

1. Cache savings aren't universal. They apply to long sessions, fixed templates, and repeated review loops. One-shot short requests gain nothing. 2. Max effort erases the savings. Independent testing shows ~60% more output tokens and per-task cost flat with Opus 5. The 40% figure is a "typical" value, not a "maxed-out" one. 3. Rising evaluation-awareness. Anthropic itself disclosed the model often recognizes being tested—a potential distribution shift between "knows I'm watched" and "assumes it isn't." 4. Capability and price decouple. The old rule "higher capability, higher price" had Fable 5.1 as its endpoint. Opus 5.5 bends that curve for the first time.

Who to believe

Anthropic's self-reported numbers answer "did the flagship get cheaper?"; independent sources (Artificial Analysis, Top10.dev, lmmarketcap, LMSPedia) answer "how much do you actually save on real tasks?" The consistently cross-confirmed figures: $4 input / $20 output / $0.20 cache read. The two conflicting numbers—Anthropic's 40% net cost drop vs AA's "flat at max effort vs Opus 5"—aren't contradictory; they measure different regimes.

---

Sources: Anthropic official release page (Sept 22), Anthropic Pricing Documentation (verified Sept 24), TechRepublic (Sept 24), Complete AI Training (Sept 25), LMSPedia evaluation page, LMMarketCap signal board (Score 95, Rank 5/438), Artificial Analysis Intelligence Index v4.3.2, Llmcostlab.com pricing archive, OpenAI Codex CLI release notes (Sept 22).

Tags

#claude-opus-5-5#gpt-6-sol#ai-coding#anthropic#openai#codex-cli#llm-pricing#coding-agents

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178635192