Rumors: Anthropic secretly testing Claude Opus 5.5, launch as early as Tuesday

Multiple leakers and media roundups claim Anthropic is secretly testing a new model under codename claude-wafer-eap, identified as claude-opus-5-5 (Claude Opus 5.5), skipping versions 5.1/5.2 and launching as early as Tuesday, in a head-on clash with OpenAI's release cadence. On Sept 22, Reddit users spotted疑似 Opus 5.5 text on Anthropic's website, raising community confidence in a Tuesday launch. Anthropic has not confirmed anything; all claims in this cluster are third-party leaks.

Unconfirmed

• Timing: @airesearch12 and @量子位 say Tuesday; @markk citing X user lyraxana says Monday or Tuesday, with Anthropic internally marking it "imminent" and partners given only 24-hour access; @CurieuxExplorer relaying @testingcatalog says this week at the earliest; @danielmac8, @scaling01 and @haider1 also say within this week amid the Anthropic–OpenAI race
• Website trace: @ResultBackground2450 reports Reddit users found疑似 Opus 5.5 wording on Anthropic's site, with screenshots circulating
• Launch prep: @applesjimmy relays netizen M1Astra's claim of early training-prep traces for Opus 5.5, suggesting an imminent launch
• Naming and skip: @新智元 and @量子位 say it skips 5.1/5.2 (or 5.2) directly
• Pricing: @xeophon lists rumored prices of $4/M input, $20/M output, $5 cache write, $0.2 cache read, calling the cheap cache reads a major win since they dominate real token spend; @量子位 says roughly 60% (elsewhere 80%) of OpenAI Astra's price; @新智元 notes input dropping to $4/M tokens
• Performance: @danielmac8 says it reportedly beats Google's Astra at lower cost and jokes about hoping for more natural voice; @量子位 claims it beats OpenAI's GPT-6 Sol/Astra; @新智元 cites tester feedback praising problem-solving intuition, deep code refactoring and long-horizon tasks
• Ecosystem: @haider1 hopes for a cheap, fast Haiku-class model to cut coding subagent costs, still missing
• Pipeline: @CurieuxExplorer relays that the next Fable model has been A/B tested on Claude Code recently

Why it matters

• If the rumored pricing holds, ultra-cheap cache reads would significantly cut costs for high-frequency Claude users
• A version-skipping launch benchmarked against GPT-6 Astra would intensify the three-way flagship race; the website trace and training-prep signs suggest the launch is in its final stretch, and Fable's A/B testing hints at a broadening lineup. Until official word, all names, dates and prices should be treated as unverified

2026-09-20 ~ 2026-09-22 · 12 related posts

• [source] Leak: Anthropic stealth-testing Opus 5.5 as claude-wafer-eap, release tipped for Tuesday — airesearch12 · 2026-09-20
• [source] Leaked Claude Opus 5.5 Pricing Shows $0.20/Mtok Cache Read, Sparking Hype — xeophon · 2026-09-20
• Rumor: Anthropic's Opus 5.5 landing this week, said to beat Astra on price and quality — daniel_mac8 · 2026-09-21
• Leaked Claude Opus 5.5 skips to testing with $4/M input tokens, undercutting Opus 5 by ~20% — 新智元 · 2026-09-21
• [source] Leaked: Anthropic to skip versions and launch Claude Opus 5.5 this Tuesday, undercutting GPT-6 Astra — 量子位 · 2026-09-21
• Leak: Anthropic Opus 5.5 may land Monday or Tuesday, with always-on adaptive thinking — mark_k · 2026-09-21
• Rumor: Anthropic May Ship Opus 5.5 This Week, Clashing With OpenAI's Tuesday Tease — CurieuxExplorer · 2026-09-21
• Claude Opus 5.5 early preparations spotted again, release looks imminent — apples_jimmy · 2026-09-22
• Opus 5.5 from Anthropic rumored to launch this week — scaling01 · 2026-09-22
• Claude Opus 5.5 Spotted on Anthropic's Website, Tuesday Release Looking Likely — ResultBackground2450 · 2026-09-22
• Claude Opus 5.5 rumored to launch this week as users await a cheap Haiku coding model — haider1 · 2026-09-22

1 near-duplicate retellings: scaling01