Tibo vs Boris: Claude Proxy Ban, Unblock, and a Codex Limit Reset

Same orange crab. Different model. Then the account lock.

On August 9, 2026, OpenAI Codex lead Tibo Sottiaux and Anthropic Claude Code lead Boris Cherny turned a developer’s Anthropic suspension into the clearest public statement yet on multi-model harnesses. Developer Alex Getman reproduced Tibo’s July claudex / CLIProxyAPI recipe almost exactly — unmodified Claude Code CLI, localhost proxy, GPT-5.6 Sol behind it — then got banned for “suspicious signals.” Public write-up: github.com/alexgetmancom/claude-proxy.

Tibo’s first reply was dry: he does not work at Anthropic, and it seemed odd to ban someone for using their harness with another model. Boris escalated in public: we don’t ban that, almost certainly a different account classifier, thanks for the ping, we are hiring. Hours later: unblocked, and a promise this should not happen to others. Tibo celebrated by resetting weekly usage for all paid ChatGPT Work and Codex users — and declined the job joke.

That is the news. The useful part is the policy: freedom of harness is now on the record from both labs, classifiers can still nuke you anyway, and a limit reset is still how OpenAI says sorry / congrats.

TL;DR — what people are asking

Timeline (public X + repo)

Boris’s hiring tweet passed 2.2M views. The joke landed; the classifier did not.

Update — same day: the celebration reset itself became the next controversy. See “The ‘performative’ backlash” below — Tibo followed up promising another reset Monday, and a chunk of replies weren’t having it.

What Getman actually ran

The claude-proxy README is not a Claude Code fork. It is a side-by-side launcher:

claudex  →  unmodified Claude Code CLI  →  127.0.0.1:8317  →  Codex / Gemini / xAI / Kimi / …
claude   →  unmodified Claude Code CLI  →  api.anthropic.com

Design choices that matter for the ban debate:

Setup path: open a fresh Claude Code session, hand it setup-prompt.md, let the agent install the proxy, browser-login the provider, verify end-to-end. Same shape as Tibo’s July “ask Claude / 5 minutes” pitch.

After the ban, Getman added a WARNING at the top of the README: he reproduced Tibo’s public post, used official unmodified Claude Code + localhost CLIProxyAPI + Sol, then got “suspicious signals.” He still does not claim the proxy was the proven cause — only the sequence.

That humility is correct. Account classifiers see login graphs, payment, device, traffic shape, prior abuse signals — not a neat “used LiteLLM = ban” bit. Boris’s public theory matches: a different classifier fired, then a human (or a fast internal path after X escalation) reversed it.

What Boris committed to (policy, not vibes)

Three sentences from Claude Code’s lead that teams will paste into Slack:

1. “We don’t ban people for using harnesses with other models.”
2. Almost certainly a different account classifier — looking into it.
3. Using our harnesses with other models is supported — proxy like LiteLLM, see docs, or ask Claude to set it up.

He added the engineering caveat the meme layer skipped: model-specific tool design and prompting is a large fraction of harness quality. Pointing Sol at Claude Code’s tool schema is not the same as native Codex, and pointing Fable at a thin OpenAI-compatible wrapper is not native Claude Code. That is the same “harness matters” thesis as Codex vs Claude Code and agent harness engineering.

Supported ≠ risk-free. Supported means product intent. Classifiers are a separate stack. Until Boris’s “doesn’t happen again” lands in production, treat new proxy traffic from a young Anthropic account as something that can trip fraud / ToS heuristics.

What Tibo did with the moment

Tibo’s celebration post:

That’s right, GPT-5.6 Sol is awesome and can be used pretty much anywhere, including in the CC harness.

To celebrate this, together with the fact that I’m not going anywhere… I have reset usage limits for all paid users of ChatGPT Work and Codex.

Then, to Boris:

Freedom of harness is important and we should let our users decide on which model is best for them.

And he is not switching: loves the team, excited about the next couple of weeks of releases.

This is the third genre of Tibo reset explainx.ai has tracked: launch-week Sol panic (5-hour cap / weekly top-ups), mid-July efficiency + 5h return (~18% longer), and now a competitive-narrative reset — “Sol belongs in any harness, including Anthropic’s.” Check your Work/Codex usage UI; these are manual top-ups, not a plan SKU change. Same pattern as the Claude / Codex / Cursor reset scoreboard.

July Tibo line that aged into a punchline: “If this gets blocked, I owe you a reset.” Anthropic blocked a user; Tibo still paid the reset — to his subscribers.

The "performative" backlash

The reset itself didn't land as cleanly as the harness-freedom quote. Two things collided:

1. Timing. Several replies pointed out the regular weekly Work/Codex limit had already renewed the day before (Aug 8), so an extra reset on Aug 9 read as free marketing rather than real headroom — Rumph: “This is just performative at this point. The weekly reset was yesterday.”
2. Tibo's own word choice. His follow-up — “I'll do another performative reset on Monday” — used “performative” himself, which read to some as self-aware humor and to others as confirmation that the resets are more about narrative timing than genuine extra capacity.

The replies split roughly three ways:

• Fatigue with the format: Shayan asked OpenAI to stop being “clever with the resets” and just ship a banked reset users can rely on, arguing the surprise-drop pattern itself was “making people angry” — a request for a durable quota mechanism instead of a recurring PR moment.
• Déjà vu with Anthropic: Shonn Li drew the direct parallel to Anthropic doing a similar “reset right after the weekly limit renewed” move a few months earlier, framing it as the same playbook regardless of which lab runs it.
• Take the free tokens and ask questions later: shiv canceled weekend plans to burn usage; Diego Carlino asked outright whether it was bait or safe to spam all weekend; others (Melos, ash) said they were now worried about burning 20x usage in anticipation of a Monday reset that might not materialize on schedule.

None of this contradicts the policy news from earlier in the thread — Boris's “we don't ban other-model harnesses” statement stands on its own. It does mean the reset that was supposed to be the celebratory footnote became a second, separate story about quota mechanics and trust, not model freedom.

What people argued about (usefully)

“Anthropic would rather hire haters than allow other models”

Teja’s dunk missed Boris’s literal sentence. The hiring joke + LiteLLM pointer is the opposite of a harness lock-in manifesto. The real failure is ops: a classifier that looks like a ban for doing the documented thing.

Hermes / “sue the harness creators”

Teknium’s jab about Anthropic subs in Hermes Agent is a different docket — subscription portability vs third-party hosts, not Getman’s localhost proxy. See Hermes vs OpenClaw. Don’t collapse “account classifier false positive” into “Anthropic sues every wrapper.”

“Just give Claude’s reset button to Tibo”

Community running gag: Tibo is the reset button. Anthropic’s equivalent is quieter (Fable window extensions, Code weekly bumps). Developers now expect public engineer accounts to fix quota and bans faster than support forms. That is unhealthy as a process and extremely effective as marketing.

Should you use LiteLLM vs claude-proxy vs official Codex plugin?

For dual-stack shops, the boring answer remains: pay both, keep claude native, use a named alias for proxy sessions, and prefer the path Anthropic’s lead just named in public.

Practical checklist (after this incident)

Multi-model Claude Code — Aug 9 hygiene
□ Read Boris: other-model harness use is supported (LiteLLM / docs)
□ Do not assume every GitHub “claudex” clone is ToS-safe
□ Keep native `claude` on api.anthropic.com
□ Proxy on 127.0.0.1 only; no exposed relay
□ Fresh / low-history Anthropic accounts: expect more classifier sensitivity
□ If suspended: appeal + wait; Getman was restored after public escalation
□ Check ChatGPT Work / Codex weekly bar — Tibo reset paid users
□ Remember: harness prompts/tools are model-specific — Sol in CC ≠ Codex native

Prompt if you want Claude to set up the documented path (not a random wrapper):

Set up Claude Code to call GPT-5.6 Sol via a localhost LiteLLM (or CLIProxyAPI)
proxy following Anthropic's current docs for third-party models.
Constraints: do not patch the Claude Code binary; do not export ANTHROPIC_BASE_URL
globally; leave the default `claude` command on Anthropic; verify with a one-line
prompt that reports which model served the reply. Explain account-classifier risk
from the Aug 9 Getman incident before changing anything.

How this fits the harness war

July was Theo + Tibo: Sol feels better in Claude Code’s crab than in Codex for some design/subagent work (full setup guide). August 9 is Tibo + Boris: that mix is allowed, classifiers can still false-positive, and OpenAI will print a reset to keep the story about user choice.

For loop engineering teams, the lesson is not “swap models daily.” It is: pin harness + model + proxy + account as a unit, instrument which endpoint you hit, and have a rollback to native Claude when fraud systems twitch. Same discipline as DAG / Planner-Worker-Critic harnesses — observability before clever routing.

Boris’s maturity ladder (Steps of AI Adoption) still applies: a shop that cannot survive one false suspension is not Step 3. Dual-subscribe and dual-login so a single classifier cannot freeze the sprint.

Honest limitations

• One public incident plus an unblock is not a published Anthropic postmortem.
• We do not have the classifier name, feature list, or whether payment/device history mattered more than the proxy.
• Getman’s sample (RAT-hunting Fable stories, etc.) is unrelated — don’t mix malware anecdotes into this ToS thread.
• Tibo resets are temporary quota, not a new Plus/Pro SKU.
• LiteLLM “supported” can still drift from docs; verify Anthropic’s live page.
• Desktop Claude Code still ignores this proxy pattern per Getman’s README.
• Star count on claude-proxy was tiny at write-up (~13) — this is a policy story, not a viral installer.

Closing

Anthropic does not ban Claude Code + Sol — a classifier did, briefly, and Boris walked it back in public. Tibo reset Work/Codex to underline freedom of harness and to decline a very on-brand recruiting joke. Developers won a policy quote and a quota top-up. They did not win a guarantee that the next community proxy will look innocent to fraud models.

Read the July claudex guide for the mechanics. Treat this post as the Aug 9 risk addendum. Follow @explainx_ai when Anthropic publishes the actual classifier fix — or when the next two weeks of OpenAI releases land.

Related on explainx.ai

• GPT-5.6 Sol in Claude Code — Claudex / CLIProxyAPI setup
• OpenAI Codex plugin for Claude Code
• Codex vs Claude Code
• Tibo ChatGPT 5-hour limit / weekly resets
• Claude / Codex / Cursor usage-limit scoreboard
• Boris Cherny — Steps of AI Adoption
• What is an agent harness?
• Hermes Agent vs OpenClaw
• ChatGPT Work vs Codex

Sources

• alexgetmancom/claude-proxy — README warning + CLIProxyAPI architecture
• Tibo Sottiaux (X, Aug 9, 2026) — odd-to-ban question; Sol-in-CC + paid Work/Codex reset; harness freedom
• Boris Cherny (X, Aug 9, 2026) — hiring joke; no ban for other-model harnesses; LiteLLM; unblock
• July 12 Tibo claudex recipe — the setup Getman followed
• CLIProxyAPI

Account status, classifier language, and quota resets reflect public X posts and the claude-proxy README as of August 9, 2026. Verify your own Anthropic and OpenAI usage dashboards and current Anthropic docs before changing production aliases. Not affiliated with Anthropic or OpenAI.