Anthropic Just Killed Opus 4.1 — Your Agents Are Broken and Nobody Told You

Related articles

• Claude Sonnet 5: The New Default Model in Claude Code
• Claude Opus 5 Arrives: Near-Fable Intelligence at Half the Price
• Beware: Claude Code v2.1.214 Quietly Closed Six Permission Holes

On August 5, 2026, Anthropic retired claude-opus-4-1-20250805 from the Claude API. Every request to that model ID now returns an error. If your coding agents, CI pipelines, or background scripts reference Opus 4.1 anywhere — in a model string, a config file, a hardcoded environment variable — they broke that day. They are still broken. And unless you specifically check, you might not know it.

This is the kind of silent failure that does not ring alarms. Your agent runs, hits the API, gets back a 404 or a redirect, and either crashes silently or — worse — continues with a fallback model that behaves differently than expected. If you are seeing sudden quality regressions, unexpected refusals, or agents that “work but feel wrong,” this is the first thing to check.

What happened

Anthropic announced the deprecation of Opus 4.1 on June 5, 2026 — exactly 60 days before retirement. The recommended replacement was claude-opus-4-8. After August 5, requests to the old model ID fail with an immediate error.

The timeline was clear if you were reading release notes:

That is a generous window by AI-industry standards. Most model providers give 30 days or less. The problem is not the warning — it is that nobody reads deprecation emails until things break.

Who is affected

You are affected if any of the following reference claude-opus-4-1-20250805:

• Claude Code model override — if you pinned the Opus model to 4.1 via --model flag or config
• API scripts — any curl call, Python script, or Node integration that sends "model": "claude-opus-4-1-20250805" in the request body
• CI/CD pipelines — GitHub Actions, GitLab CI, or any automated workflow that calls the Anthropic API with the old model string
• Agent harnesses — Codex, Hermes, Goose, or any framework that lets you configure the model endpoint
• Bedrock / Vertex deployments — if you provisioned Amazon Bedrock or Google Vertex with the Opus 4.1 model ID, those integrations break too
• Third-party tools — any SaaS product that internally routes to the Anthropic API and cached the old model identifier

The failure mode depends on your setup. Direct API calls return an HTTP error immediately. Framework-level integrations may silently fall back to a different model, which changes behavior without obvious failure signals. A Bedrock deployment that was routing to Opus 4.1 may show up as “model not available” in CloudTrail.

How to diagnose

The fastest check is a direct API call with the old model ID:

curl -s https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-4-1-20250805",
    "max_tokens": 64,
    "messages": [{"role": "user", "content": "Reply with OK"}]
  }'

If this returns an error — model not found, invalid model ID, or similar — your integration is broken.

For code-level checks, search your entire workspace:

# Find any reference to the old model ID
grep -r "opus-4-1" . --include="*.json" --include="*.yaml" --include="*.yml" --include="*.env" --include="*.toml" --include="*.cjs" --include="*.mjs" --include="*.py"

# Also check for the older identifier patterns
grep -r "claude-opus-4-1-20250805" . --include="*.json" --include="*.yaml" --include="*.env" --include="*.py" --include="*.js"

If your agent harness has a model configuration file (often .claude/settings.json, agent.config.json, or an environment variable like ANTHROPIC_MODEL), check it directly.

What to migrate to

Anthropic recommends claude-opus-4-8 as the drop-in replacement. In practice, there are three viable options depending on your use case and budget:

For most coding agent workflows, Opus 5 or Sonnet 5 are better choices than Opus 4.8. Opus 5 launched July 24 and is the current state-of-the-art for agentic coding. Sonnet 5 delivers close-to-Opus quality at roughly one-third the cost. Only reach for Opus 4.8 if you have a specific reason — like needing the Priority Tier service level, which Sonnet 5 does not support.

If you were using Opus 4.1 through Claude Code (not the API directly), Claude Code already auto-upgrades to Opus 5 as the default Opus model. You likely do not need to change anything — unless you had explicitly pinned the model version in a config override.

The hidden gotcha: temperature and top_p now return 400

Even if you have already migrated to Opus 4.8 or 5, there is a breaking behavioral change that catches experienced developers off guard.

In Opus 4.7 and 4.8 (and Opus 5), passing non-default values for temperature, top_p, or top_k returns HTTP 400. These parameters are no longer user-configurable on the flagship Opus models. The only way to steer model behavior is through prompting.

This matters because a common pattern in agent pipelines is:

# This worked on Opus 4.1 — it BREAKS on 4.8 / Opus 5
response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=4096,
    temperature=0.3,  # ← HTTP 400
    top_p=0.9,        # ← HTTP 400
    messages=[...]
)

If your agent code sets these parameters, update your code to remove them. Use prompting strategies instead — system prompt instructions like “be precise and conservative” or “explore multiple approaches” achieve similar effects without the API-level parameter.

Sonnet 5 does not have this restriction. If you need user-configurable temperature, Sonnet 5 still supports it.

The broader lesson: model lifecycle is an operational dependency

The Opus 4.1 retirement is not an isolated event. Anthropic retired Opus 4 and Opus 4.1 simultaneously. They gave 60 days. That is fair. But the pattern of model retirements breaking production is becoming predictable.

Every coding agent operator should have:

1. Model ID as an explicit config variable, not hardcoded in source. Set it in an environment variable or config file that is easy to update in one place.
2. A deprecation monitoring practice. Subscribe to the Anthropic changelog, OpenAI release notes, and the model deprecation pages. Check monthly.
3. Error-rate alerts. If your agent pipeline starts returning 400s or model-not-found errors at a rate above your baseline, you need to know immediately — not when a developer notices the agent “feels slower.”
4. Fallback model strategy. If your primary model goes down, what does the agent fall back to? If the fallback is “silent quality regression,” you have a monitoring gap.

The AI model market is moving fast. Models ship, deprecate, and get replaced in weeks. Treating the model layer as a stable API is a mistake. Treat it like a dependency that needs its own lifecycle management.

Quick migration checklist

# 1. Find old model references
grep -r "opus-4-1" . --include="*.json" --include="*.yaml" --include="*.env" --include="*.py" --include="*.js"

# 2. Replace with your chosen model
#    Option A: Opus 4.8 (direct replacement)
#    Option B: Opus 5 (latest, best for coding)
#    Option C: Sonnet 5 (cheapest, near-Opus quality)

# 3. Remove temperature/top_p/top_k from Opus 4.7+ calls
grep -r "temperature" . --include="*.py" --include="*.js" --include="*.ts" | grep -i opus
grep -r "top_p" . --include="*.py" --include="*.js" --include="*.ts" | grep -i opus
grep -r "top_k" . --include="*.py" --include="*.js" --include="*.ts" | grep -i opus

# 4. Test with a direct API call
curl -s https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "max_tokens": 64,
    "messages": [{"role": "user", "content": "Reply with OK"}]
  }'

# 5. Verify your agent runs end-to-end

Bottom line

Claude Opus 4.1 is gone. The 60-day grace period is over. If your agents were pinned to the old model ID, they are returning errors right now — or worse, silently falling back to a different model with different behavior. Check your configs, update the model string, and while you are at it, remove any temperature or top_p parameters from Opus-family API calls before they become your next 400 error.

The model layer is not infrastructure you set and forget. It is a living dependency that needs active lifecycle management. Treat it accordingly.