🚨 Breaking — April 4, 2026

AI Model Replacement Guide

Prepared by Dwight · April 4, 2026 · Anthropic killed OpenClaw OAuth today

Table of Contents

  1. The Situation
  2. Sonnet Replacement
  3. Opus Replacement
  4. Full Model Rankings — April 2026
  5. Subscription OAuth Options
  6. Can GPT-5.4 Do What Sonnet Does?
  7. Transition Plan (3 Options)
  8. Local Model Option (Gemma 4)
  9. Final Recommendation

The Situation

Anthropic killed third-party OAuth today (April 4, 2026 at 12pm PT). You can no longer use a Claude subscription (Pro/Max) to power OpenClaw. Two options remain for Claude:

  • Anthropic API key — pay-per-token ($3/$15 per MTok for Sonnet, $5/$25 for Opus)
  • Switch to a different provider entirely

This guide answers: What should you actually do?

Section 1: The Sonnet Replacement

Current role: Fergus's main agent — everyday conversation, task management, delegation, Discord responses, voice message handling, tool calls, persona consistency, multi-agent orchestration.

🥇GPT-5.4 via OpenAI Codex OAuth — CAN REPLACE SONNET

This is the move. Here's why:

  • Quality: AI Top 40 score of 100.0 (vs Sonnet's #5-6 range). Terminal-Bench 2.0: 75.1% vs Sonnet's 59.1%.
  • Tool calling: Has "Tool Search" — dedicated multi-tool orchestration. "Slight edge in complex orchestration where orchestrator must track multiple simultaneous sub-task states." This is exactly what Fergus does.
  • Session memory: ZooClaw benchmark rated GPT-5.4 "Excellent" for memory reliability after compaction — Sonnet rated "Good".
  • OpenClaw community: "5.4 is more than capable of running main agent 95% of the time." Named ZooClaw overall winner above Sonnet AND Opus.
  • Access: Codex OAuth via ChatGPT Plus ($20/mo) or Pro ($200/mo). The ONLY subscription-based OAuth that still works in OpenClaw for a top-tier model.

Where GPT-5.4 is WORSE than Sonnet

  • Speed: 20-30 tok/s vs Sonnet's 44-63 tok/s. Slower Discord responses — biggest real-world tradeoff.
  • Persona compliance: GPT-5.4 wants to DO work. Sonnet follows "don't do work, delegate" rules more naturally. Needs stronger guardrails.
  • Discord formatting: Occasionally generates markdown tables (broken in Discord). Add explicit rules to SOUL.md.
  • Verbosity: Tends toward longer responses. Add "be concise" rules.
Verdict: GPT-5.4 CAN replace Sonnet. Stronger on tool calling, session memory, and multi-step orchestration. Weaker on speed, persona compliance, and formatting — all fixable with prompt engineering. Transition will require 1-2 weeks of SOUL.md tuning.
🥈Claude Sonnet 4.6 via Anthropic API Key

Keep Sonnet by moving to API billing. Identical quality. Cost: $3/$15 per MTok (~$50-150/mo with 90% prompt caching discount for heavy daily use).

Verdict: Obviously works — it's the same model. Question is whether $50-150/mo variable beats $20/mo flat for GPT-5.4 Plus, which is arguably better.
🥉Gemini 3.1 Pro via Google API Key — CANNOT FULLY REPLACE

Strong benchmarks (GPQA 94.3%, SWE-bench 80.6%) but no subscription OAuth in OpenClaw, Google is suspending accounts that use their API heavily with OpenClaw agents, and persona consistency is "underwhelming out of the box."

Verdict: CANNOT replace Sonnet for this use case. Too risky, no subscription, weak persona handling.
Sonnet Replacement Recommendation:
  • Primary: GPT-5.4 via Codex OAuth (ChatGPT Plus $20/mo or Pro $200/mo)
  • Fallback: Keep Sonnet via Anthropic API key ($50-150/mo variable)

Section 2: The Opus Replacement

Current role: Dwight — deep analysis, strategy, research, complex reasoning, comparison tasks, financial analysis, product research.

🥇GPT-5.4 via OpenAI Codex OAuth — CAN REPLACE OPUS (for 80% of tasks)
  • AI Top 40 #1 (100.0) vs Opus at #2 (93.2). Leads on 8 of 10 benchmarks.
  • Matches Opus on: multi-step reasoning, knowledge work, structured analysis, research synthesis.
  • Falls short: Opus has a unique quality for nuanced, creative reasoning — "thinking around corners." GPT-5.4 is more systematic, Opus more intuitive. For the hardest 20% of Dwight tasks, Opus still has an edge.
Verdict: GPT-5.4 CAN replace Opus for most Dwight tasks. For the hardest 20%, you lose some nuance.
🥈Claude Opus 4.6 via Anthropic API Key — STAYS AS OPUS

Arena #1 (Elo 1,504). Arena Code #1 (Elo 1,548). The acknowledged quality ceiling. Cost: $5/$25 per MTok. Dwight is used selectively so realistic monthly: $30-80.

Verdict: Opus via API CAN stay as Opus. $30-80/mo for the best reasoning model is reasonable for selective use.
Opus Replacement Recommendation:
  • Primary: Keep Opus via Anthropic API key ($30-80/mo) — cheap enough for selective Dwight use
  • Alternative: Use GPT-5.4 for everything (accepts the 20% quality loss on deepest reasoning)

Section 3: Full Model Rankings — April 2026

Ranked by overall capability. Pricing, access method, and value rating for each.

Tier S — Frontier (Best Models Alive)

#1
GPT-5.4 OpenAI
AI Top 40: 100.0 · Terminal-Bench: 75.1% · SWE-bench: ~80%
API: $2.50/$15 per MTok · ✅ Codex OAuth subscription ($20-200/mo)
Best all-around agentic model. Tool calling. Computer use. Multi-step execution.
🔵 $20/mo
#2
Claude Opus 4.6 Anthropic
Arena: #1 (1,504) · Arena Code: #1 (1,548) · SWE-bench: 79.6%
API: $5/$25 per MTok · ⚠️ API key only (OAuth killed April 4)
Best human-preference quality. Best coding. Best nuanced reasoning. Best persona adherence.
🟡 Fair
#3
Grok 4 xAI
AI Top 40: 86.6 · HLE: 50.7% (#1) · Strong reasoning
API: $3/$15 per MTok · ✅ API key only. No OAuth.
Strongest on Humanity's Last Exam. Good deep reasoning.
🟡 Fair
#4
Gemini 3.1 Pro Google
GPQA: 94.3% (#1) · SWE-bench: 80.6% · ARC-AGI-2: 77.1% (#1)
API: $2/$12 per MTok · ✅ API key. ❌ No OAuth subscription.
⚠️ Risk Google suspending accounts using API with OpenClaw agents.
🟢 Great value

Tier A — Near-Frontier

#5
Claude Sonnet 4.6 Anthropic
SWE-bench: 79.6% · 44-63 tok/s · Excellent instruction following
API: $3/$15 per MTok (90% cache discount) · ⚠️ API key only (OAuth killed April 4)
Fastest frontier-class model. Best instruction following. Best persona consistency. Best Discord formatting.
🟢 Excellent
#6
GPT-5.4 Pro OpenAI
Enhanced reasoning version
API: $30/$180 per MTok (12x standard) · ✅ API key. Not on Codex OAuth.
Maximum reasoning when cost is no object.
🔴 Poor value
#7
GPT-5.2 OpenAI
Strong general-purpose. "Most reliable instruction following" per OpenAI.
API: $1.75/$14 per MTok · ✅ API key.
🟢 Excellent
#8
Qwen3-Max Alibaba
API: $1.20/$6 per MTok · Or Alibaba Coding Plan: $10-50/mo flat
✅ API key + ✅ Alibaba Coding Plan OAuth
⚠️ Quality Significantly below Sonnet/GPT-5.4. "TV robot" persona compliance.
🟢 Budget

Tier B — Strong Performers

#9
Grok 4.2 xAI
Fast variant. API: $2/$6 per MTok · ✅ API key. 2M context window.
🟢 Excellent
#10
GPT-5.3-Codex OpenAI
400K context. Strong coding specialist. API: $1.75/$14 per MTok
✅ Codex OAuth (subscription) — included in Plus/Pro. This is what Cody/Bomb use.
🔵 Included
#11
GPT-5.3-Codex-Spark OpenAI
1,000+ tok/s on Cerebras. Ultra-fast coding. Pro subscription only.
✅ Codex OAuth · Separate usage pool from GPT-5.4 (different hardware)
Pro users get GPT-5.4 quota PLUS Spark quota independently — more total headroom.
Pro only
#12
DeepSeek V3.2 DeepSeek
API: $0.28/$0.42 per MTok (27x cheaper than Opus)
⚠️ Privacy Chinese company. Data handling concerns for business use.
🟢 Cheapest
#13
Mistral Large 3 Mistral
API: $0.50/$1.50 per MTok · European data handling.
🟢 Excellent

Tier C — Budget / Specialist

#15
Gemma 4 31B Google Open
AA Intelligence Index: #2 among open models (score 39). MMLU-Pro 85.2%. Native function calling.
✅ Local via Ollama · Free if you have hardware. Apache 2.0 license.
⚠️ Gap Sonnet has a 43-point advantage on AA Intelligence Index. Good for local fallback, not primary agent.
Free local
#16
Llama 4 Maverick Meta Open
AI Top 40: #31. 400B total params, MoE. Multimodal.
Free locally · API via OpenRouter/Groq · CANNOT replace Sonnet.
Fair

Section 4: Subscription OAuth in OpenClaw — Full List

Which providers let you pay a flat monthly rate and use it through OpenClaw?

✅ Confirmed Working Subscription OAuth

Provider Subscription Price Models Quality for Main Agent?
OpenAI Codex ChatGPT Plus $20/mo GPT-5.4, GPT-5.3-Codex, GPT-5.4-mini YES — Top tier
OpenAI Codex ChatGPT Pro $200/mo All Plus + Spark, 6x limits YES — Top tier + headroom
Alibaba Coding Plan Lite $10/mo ($3 first mo) Qwen3.5+, Kimi K2.5, GLM-5, MiniMax M2.5 No — budget tier
Alibaba Coding Plan Pro $50/mo ($15 first mo) Same models, 5x more requests No — budget tier
MiniMax Coding Plan OAuth ~$10-20/mo MiniMax models No — budget tier
Z.AI / GLM Coding Plan OAuth ~$10-20/mo GLM models No — budget tier

❌ No Subscription OAuth

The reality: OpenAI is the ONLY provider offering top-tier subscription OAuth in OpenClaw.

Section 5: Can GPT-5.4 Do What Brent Does With Sonnet?

✅
Long complex sessions without degrading
ZooClaw benchmark: "Excellent" for memory reliability after compaction — same as Opus, BETTER than Sonnet ("Good").
⚠️
Follow SOUL.md / AGENTS.md persona instructions
YES, WITH WORK. More opinionated than Sonnet. Needs stronger delegation rules, explicit "no markdown tables," explicit "be concise." Set personalityOverlay: "off".
✅
Handle tool calls reliably
BETTER than Sonnet. Terminal-Bench 75.1% vs 59.1%. Designed for agentic tool use.
✅
Multi-agent orchestration
"Instruction-following precision gives it a slight edge in complex orchestration where orchestrator must track multiple simultaneous sub-task states."
⚠️
Format Discord messages properly
YES, WITH RULES. Occasionally generates markdown tables, tends toward longer responses. Fixable with explicit SOUL.md rules.
✅
Flat-rate subscription
The ONLY top-tier model with subscription OAuth in OpenClaw right now.
Bottom line: GPT-5.4 CAN do everything Brent does with Sonnet. It's actually BETTER at tool calling and session memory. Worse at delegation compliance, formatting, and speed — all fixable with prompt engineering. Will take 1-2 weeks of SOUL.md tuning.

Section 6: Transition Plan

Option A: All-In OpenAI (Simplest, Cheapest)
$200/mo (Pro) or $20/mo (Plus)
  • Main Agent (Fergus): openai-codex/gpt-5.4
  • Sub-Agent Dwight: openai-codex/gpt-5.4
  • Sub-Agent Scout: openai-codex/gpt-5.4-mini
  • Coding (Cody/Bomb): openai-codex/gpt-5.4 (already on this)

Pros

  • One provider, one bill, flat rate
  • Generous limits

Cons

  • No Opus for hardest reasoning
  • Slower speed
  • Prompt tuning needed

SOUL.md additions needed:

## 🚨 GPT-5.4 SPECIFIC RULES
- NEVER use markdown tables. Use bullet lists instead. Always.
- Keep responses concise. Don't over-explain.
- You are a DISPATCHER. Do NOT do work yourself. This model
  has a tendency to try to do things — resist it. Delegate EVERYTHING.
- When formatting for Discord: use bold headers + bullet lists.
  No tables. No code blocks for non-code content.

openclaw.json config:

{
  "plugins": {
    "entries": {
      "openai": {
        "config": {
          "personalityOverlay": "off"
        }
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "openai-codex/gpt-5.4"
      }
    }
  }
}
Option B: Hybrid (Best Quality, Higher Cost)
~$280-430/mo total
  • Main Agent (Fergus): openai-codex/gpt-5.4 ($200/mo Pro or $20/mo Plus)
  • Sub-Agent Dwight: anthropic/claude-opus-4-6 (API key, $30-80/mo)
  • Sub-Agent Scout: openai-codex/gpt-5.4-mini (included)
  • Coding (Cody/Bomb): openai-codex/gpt-5.4 (included)
  • Fallback: anthropic/claude-sonnet-4-6 (API key, if GPT-5.4 hits limits)

Pros

  • Best of both worlds
  • Opus for hard problems
  • GPT-5.4 for everything else

Cons

  • Two providers, two billing methods
  • More complexity
Option C: Stay on Anthropic API (Minimal Change)
~$100-250/mo variable
  • Main Agent (Fergus): anthropic/claude-sonnet-4-6 (API key)
  • Sub-Agent Dwight: anthropic/claude-opus-4-6 (API key)
  • Coding (Cody/Bomb): openai-codex/gpt-5.4 (Codex OAuth, Plus $20/mo)

Just swap the Anthropic auth from OAuth token to API key. Same SOUL.md works. Variable monthly cost — could spike.

Section 7: Local Model Option — Gemma 4

Can Gemma 4 replace Sonnet if run locally?

Short answer: No.

  • AA Intelligence Index: Gemma 4 scores 39 vs Sonnet's 82+ (43-point gap)
  • Gemma 4 is the BEST open/local model — but "best local" is still a tier below frontier cloud models.

Hardware Requirements

Best Local Models for OpenClaw — April 2026

  1. Gemma 4 31B — Best overall. Native function calling. Apache 2.0. Needs 32GB+ RAM.
  2. Qwen3.5 27B — "Matches GPT-5 Mini." Good tool calling. Needs 32GB+ RAM.
  3. Gemma 4 26B MoE — Good for its effective size (3.8B active). Best "bang for buck" — can squeeze into 16GB.
  4. DeepSeek-Coder small variants — Good for local coding agents.
Recommendation: Don't buy hardware for local models as a Sonnet replacement. The quality gap is too large. Local models are great as emergency fallback, privacy-sensitive tasks, or saving money on low-priority Scout-level work. But for the main Fergus agent? Cloud models are the answer.

Final Recommendation — What To Do Right Now

Immediate (Today)

  1. Set up Anthropic API key in OpenClaw config — keeps everything working while you transition. Cost: ~$50-150/mo for Sonnet, $30-80/mo for selective Opus.
  2. Keep Codex OAuth for Cody/Bomb (already working on Plus $20/mo).

This Week

  1. Test GPT-5.4 as main agent on a non-critical channel. Try it in #fergus for a day.
  2. Tune SOUL.md for GPT-5.4 quirks (anti-table rules, stronger delegation language, conciseness).

Within 2 Weeks

  1. If GPT-5.4 works → upgrade to Pro ($200/mo) and go all-in (Option A or B).
  2. If GPT-5.4 doesn't match Sonnet's persona → stay on Anthropic API (Option C), accept variable cost.

The Money

  • Cheapest good option GPT-5.4 Pro all-in = $200/mo flat
  • Best quality option GPT-5.4 Pro + Opus API for Dwight = $230-280/mo
  • Status quo Anthropic API for everything = $100-250/mo variable + Plus $20/mo for Codex
The bottom line: GPT-5.4 is the replacement. It's ranked #1 overall. The OpenClaw community endorses it. It's the only top-tier model with subscription access. The only question is whether you can tune your prompts to match Sonnet's persona compliance — and based on everything reviewed here, you can. It just takes a week or two of iteration.

AI Model Replacement Guide · Prepared by Dwight · April 4, 2026 · For Brent's OpenClaw stack