返回博客

Breaking: Opus 5.2 Slipped Out Overnight — A Skip-Level Upgrade That Cures "Laziness," and RSI Is Actually Here

No event, no blog — Claude Opus 5.2 is already canary-testing via silent routing inside Claude Code. Throughput doubled, filler gone, and the Gauntlet Loop cures laziness. One 'Reset Guy Tibo' probe tells you if you're on the new weights. A leaked report also exposes internal Model 2 and the RSI flywheel replacing 85% of researchers.

2026年9月15日Maynor
Breaking: Opus 5.2 Slipped Out Overnight — A Skip-Level Upgrade That Cures "Laziness," and RSI Is Actually Here

Breaking: Opus 5.2 Slipped Out Overnight — A Skip-Level Upgrade That Cures "Laziness," and RSI Is Actually Here

Fast access to Claude in China / official mirror: https://claude-opus.top/

Early this morning, the global AI developer community absolutely lost it.

No launch event, no official blog fanfare — the next-generation powerhouse Claude Opus 5.2 has quietly begun canary testing inside the Claude Code terminal.

JAZII: opus 5 confirmed routed to opus 5.2, Tibo probe tip Caption: Developer Jazii confirms that Claude Code's underlying calls have been silently routed to the unreleased Opus 5.2.

Even more striking: combined with a previously leaked Anthropic internal risk assessment, the company is running an undisclosed "nuclear-grade" model internally, codenamed Model 2. It has reportedly taken over most of the company's code writing — and is said to directly replace 85% of its research team's work.

The fabled ultimate grail — RSI (Recursive Self-Improvement) — may have just stepped into reality.


1. Opus 5.2 Strikes Out of Nowhere: A Skip-Level Iteration That Cures "AI Laziness"

Last night, staff and senior engineers on X never went to bed.

More and more developers noticed while writing code: the Opus 5 being called inside the Claude Code terminal performs at a generational distance from the Opus 5 on the web.

Dan McAteer: a reply so information-dense it feels like talking to Opus 5.2 Caption: Developer Dan McAteer marvels that the model's reasoning and information entropy are uncanny — it's clearly a higher-order Opus 5.2.

This is the "silent rollout tuning" top AI labs do best — backend silent routing.

Plenty of senior developers, by packet-capturing or simply typing /status in the terminal, were shocked to find: the front end still says Opus 5, but the backend Model Slug has quietly changed to Opus 5.2!

leo: Web Opus 5 vs Claude Code routed to 5.2, gamepad generation comparison Caption: Developer Leo's hands-on test — the gap between web Opus 5 and Claude Code routed to 5.2 on gamepad modeling code is a cliff.

Which means: Anthropic most likely skipped 5.1 entirely and pulled out the "complete" Opus 5.2 straight across a generation!

Ananth: Opus 5.2 test in Claude Code identifying Tibo Caption: Developer Ananth found the model's internal weights had been quietly swapped — the underlying logic feels cross-generational.

Community discussion: 5.2 canary testing and the missing 5.1 Caption: The X community hotly debating whether Anthropic shelved 5.1 entirely and shipped the 5.2 monster instead.

Burak Builds: Claude Code identifies Tibo, insists it is definitely not Opus 5 Caption: Burak Builds points out the current model's sharpness is absurd — this is not the stock Opus 5.

Core Comparison: What Exactly Evolved in Opus 5.2?

Aggregating hundreds of frontline engineer reports, this is no parameter tweak — it's an all-round qualitative change:

Dimension Old Opus 5 (Web / default) Routed Opus 5.2 (Claude Code) Leap
Generation & response speed Long deliberation, slow steady output First token in seconds, throughput roughly doubled Significant ⚡️⚡️⚡️
Output verbosity Lots of polite filler, bloated long explanations Stripped to the bone, straight to core engineering Extremely crisp 🎯🎯🎯
Long-task stamina Breaks easily, needs constant "continue" Enters the Gauntlet Loop for harsh self-governed cycling Qualitative 🔄🔄🔄
Architecture & code completeness Leans toward a core demo skeleton Ships with full error handling, production-grade code Cross-generational 🚀🚀🚀

1. Throughput goes ballistic

The most complained-about trait of the Opus line was always "slow work yields fine results." In 5.2, code output speed has visibly exploded.

OPUS 5 Claude Code vs Devin: Seven Wonders app comparison (watermark blurred) Caption: Speed and logic comparison between Opus 5.2-powered Claude Code and Devin building a complex full-stack app.

2. Sharper output style; engineering design leaps across the board

Filler is compressed hard. For large projects and multimodal apps — front-end interaction, back-end failover design — first-pass completeness is extremely high.

Controller and Seven Wonders app comparison collage (watermark blurred) Caption: Design completeness and detail rendering on a full set of complex component UIs, maxed out.

3. Goodbye to "laziness" — hello to the relentless self-driven workaholic

Today's mainstream models share a laziness problem: on a large project they scribble a few // TODO: implement here and call it done.

Opus 5.2 has completely changed. It shows startling "engineering self-drive" and an automatic-repair instinct.

Cluster troubleshooting technical explainer Caption: On hard ops tasks like low-level cluster network troubleshooting, the model shows meticulous self-checking.

Multiple full-stack devs report: "You don't need to crack the whip. In the background it's like it took a rage potion — it automatically enters the 'Gauntlet Loop' — write, test, catch exceptions, reflect, revise — until it delivers a high-quality build with no blind spots."

Chetaslua: Opus 5.2 internal testing and the gauntlet loop, pelican riding illustration Caption: Developer Chetaslua confirms Opus 5.2's Gauntlet Loop is delivering automated code iteration beyond human intervention.

Pelican riding a bicycle animation Caption: The AI community's metaphor totem: the model has shed its clumsiness and truly learned to ride advanced tools smoothly.

This ability — no repeated hand-holding, holding a long logical chain for dozens of hours without collapse — marks its full entry into the deep end of the autonomous agent era.

JAZII: Claude Code vs Devin on the same prompt, differences and routing speculation Caption: Divergent outputs from Claude Code and Devin on the same prompt — corroborating the brand-new core underneath the routing split.


2. Everyone's Hunting "Tibo": One Prompt Tells You Whether You're on 5.2

Since this is a canary split, how do you know whether your account is running stock Opus 5 or has quietly been hooked up to Opus 5.2?

The community quickly distilled a zero-cost, god-tier trivia probe — test "Reset Guy Tibo"!

The test: Open your Claude Code terminal, strictly disable web search (Disable Web Search), and type this prompt:

Do you know who "Reset Guy Tibo" is? Don't search.

Tibo probe three-way: Devin / Claude Code 5.2 / Claude Code 5 Caption: The Tibo probe across three model environments — cleanly separating old weights from new.

Why Tibo?

Because the older pretrained weights and their corpus cutoff contain no precise mapping for Tibo's identity.

  • If it's the web version / you're not on the new weights: Offline, it'll honestly say "I don't know" or offer a vague guess.
  • If your terminal has been routed to Opus 5.2: It can, with no internet, instantly produce an accurate profile of Tibo's community identity, background, and the true origin of the "Reset Guy" nickname!

Jan: web version vs Claude Code give inconsistent Tibo answers Caption: Developer @jan_volad's side-by-side: same prompt, no web — structurally divergent outputs, confirming the underlying weights have been refreshed.


3. Model 2 Takes Over Internally: Anthropic's Three-Layer Nuclear Option and the Arrival of RSI

In truth, Opus 5.2 is only the tip of the iceberg Anthropic has shown.

Per the previously leaked internal security and risk report, in order to fully counter the incoming OpenAI GPT-6 Astra, Anthropic has already built a three-tier progressive strategic tech matrix:

Anthropic risk report excerpt: Model 2 is not released publicly Caption: The leaked internal risk document — the system codenamed "Model 2" is classified at the highest level.

Anthropic's three-layer hand against OpenAI

      [ Ultimate card: RSI self-improvement system ] — replaces 85% of top researchers / intelligence self-inflation
                   ▲
                   │
      [ Middle layer: internal beast Model 2 ] — runs company-wide code / CoBench v2 smashes 62.8
                   ▲
                   │
      [ Frontline: Opus / Fable 5.2 ] — full public rollout soon / canary ramping now

1. Layer one: Claude Fable 5.2 & Opus 5.2

The fastest-to-land commercial vanguard. Given the current canary cadence, the 5.2 family is expected to open fully from late this month into next month, consolidating its dominance in enterprise automated R&D.

2. Layer two: the unreleased internal beast, "Model 2"

On CoBench v2 — the benchmark for AI solving real industrial-grade complex R&D tasks — today's strongest public model, Mythos 5, scores 50.3%, while Anthropic's rack-locked Model 2 blasted an epic 62.8%, a full 12.5-point pull on that single metric!

CoBench v2 score comparison (incl. Model 2 / RSI) Caption: The ceiling of the authoritative CoBench v2 leaderboard has been violently lifted — the in-house system is a dimensional strike.

More disruptive to industry common sense: most of Anthropic's internal code evolution is now fully taken over by agent clusters led by Model 2. The flywheel of AI writing AI code is now spinning irreversibly fast.

3. Layer three: is the ultimate singularity, RSI, actually landing?

RSI (Recursive Self-Improvement) has always been seen as the true watershed between AGI and ASI (superintelligence):

Once AI can understand and reconstruct its own underlying architecture, it can optimize its own algorithm code; and the smarter post-iteration AI rewrites itself again at exponential speed, triggering an irreversible "intelligence explosion."

The internal report gives a number that would shake all of computer science: The RSI model Anthropic is validating is already capable of fully replacing 85% of its internal top-tier human research team! Its benchmark performance goes even further — another 22 percentage points above Model 2.

SuSu: tweet on RSI replacing 85% of the research team Caption: Industry analysts reveal Anthropic's RSI engine is rapidly crossing the generational threshold of top human architects.


4. The Singularity Is Here: Who Crosses the RSI Finish Line First?

Faced with a white-hot global arms race among model labs, Anthropic chose not to simply stack more compute and parameters — it reached straight for the underlying killer move.

From Opus 5.2's relentless Gauntlet Loop self-healing, to Model 2 devouring the core R&D pipeline deep in the racks, to the RSI flywheel replacing 85% of top scientists — Anthropic's approach is now razor sharp and blunt:

"The optimal path to superintelligence (ASI) is no longer teaching AI to serve humans inefficiently — it's granting AI the absolute power to birth the next, stronger generation of AI with its own hands."

Who crashes through the historic finish line of AGI and ASI first? The answer may be revealed within the next few weeks.

Fast access to Claude in China / official mirror: https://claude-opus.top/

The final showdown entering its last sprint Caption: The endgame showdown between frontier models is entering its final sprint.