Migrate GPT-4o-mini to GPT-5.6 Luna: Checklist

Sibling posts: Luna guide · alias gotcha · cost per completed task · Luna vs Haiku.

What you are trading

Mini remains one of OpenAI’s best absolute bargains. Luna’s value is current-generation GPT-5.6 at a nano-ish price after the July 30 cut — not beating mini’s $0.15/$0.60 row. Migrate when quality, context, or tool features matter more than squeezing the last tenth of a cent.

• Context: mini ~128k · Luna ~1.05M (with the usual Luna long-context recall caveats)
• Family: mini is GPT-4o-era · Luna is GPT-5.6 with Responses / programmatic tool calling
• Vision: both accept images on common configs — re-test your image prompts
• CloudAxis: Auto Mode is already Luna — hosted operators do not edit model strings

Price table (API list)

Confirm on OpenAI pricing before you lock a budget. Late July 2026 list rates we use in planning:

On a balanced 2:1 input:output mix, Luna’s output rate is what hurts. Example: 10M in + 5M out → mini ≈ $4.50 · Luna ≈ $8.00. Cache-heavy workloads narrow the input gap (Luna cached input $0.02) but do not erase the output delta. Model your mix on the cost calculator.

Migration checklist

Do these in order:

1. Inventory every gpt-4o-mini (and snapshot ids) in app config, gateways, eval harnesses, and cron
2. Replace with gpt-5.6-luna — search also for bare gpt-5.6 mistakes
3. If you send tools on Chat Completions, set reasoning_effort: "none" or move to the Responses API
4. Drop unsupported mini-only parameters; keep temperature/max tokens within Luna limits (128k max out)
5. Re-run your eval set — especially JSON schema, tool loops, and long-doc tasks
6. Compare cost per successful task (including retries), not only $/1M
7. Roll out behind a flag; watch error rates and p95 latency for a week

API and parameter gotchas

Most “migrations that broke overnight” are parameter mismatches, not model quality:

• Alias trap: gpt-5.6 ≠ Luna — it routes to Sol (flagship). Always pin gpt-5.6-luna. See the alias post.
• reasoning_effort: Chat Completions + function tools on GPT-5.6 often need "none" or the call fails / behaves oddly. Responses API is the cleaner long-term surface.
• JSON mode: re-check structured outputs. Mini sometimes fences JSON in markdown; Luna may return raw JSON — update parsers either way.
• max_tokens vs max_completion_tokens: align with your SDK’s GPT-5.x naming; do not assume 4o-mini parameter names copy cleanly.
• System prompt drift: GPT-5.6 instruction-following differs from 4o-mini. Re-tune short agent system prompts after the swap.

Measured example (why evals matter)

On 2026-07-31 we ran the same 40-SKU monitor prompt used in our cost-per-task series: extract cheapest SKU/vendor/price/spread/n as JSON. Ground truth: spread_usd = 29.

• gpt-4o-mini: $0.000105 · 502/50 tokens · 1347 ms · incorrect spread_usd (28 vs expected 29); fenced JSON
• gpt-5.6-luna: $0.000141 · 501/34 tokens · 870 ms · correct raw JSON

Mini was cheaper on that single call (~1.34× Luna) and still failed the task. Luna cost a fraction of a tenth of a cent more and passed. That is the migration heuristic: pay for completed work. Log in sites/cloudaxis/content/gpt-5-6-luna-vs-4o-mini-measurement.mjs. One task ≠ forever — re-run your eval suite.

Code-shaped diff

Typical client change (illustrative):

// before
model: "gpt-4o-mini"

// after
model: "gpt-5.6-luna",
reasoning_effort: "none"  // if using Chat Completions + tools

On CloudAxis hosted chat you do not edit this — Auto Mode is already Luna on Free and paid. The checklist above is for teams calling OpenAI (or OpenAI-compatible gateways) directly with hard-coded mini.

When to stay on mini

Keep mini when your evals are already green, output volume is huge, and you do not need GPT-5.6 features or mega-context. There is no virtue in migrating for the blog post.

• Stay on mini for high-QPS classification, simple extraction with green evals, and pure $/token floors
• Move to Luna when long-doc agents, tool loops, vision, or measured task failures make retries cost more than Luna’s list premium
• Skip bare gpt-5.6 always — that is Sol billing, not a “safe default”

Related reading

• GPT-5.6 Luna complete guide
• gpt-5.6 → Sol gotcha
• Cost per completed task
• Luna vs DeepSeek
• Luna vs Claude Haiku
• AI cost calculator

Frequently asked questions

Is GPT-5.6 Luna cheaper than GPT-4o-mini?

Usually not on list rates. Mini is about $0.15/$0.60 vs Luna’s $0.20/$1.20 per 1M. Migrate for capability and task success rate, then re-check cost per success.

What model ID replaces gpt-4o-mini?

Use gpt-5.6-luna. Do not use gpt-5.6 — that alias routes to Sol (flagship pricing).

Will my tool-calling code break?

It can on Chat Completions if you omit reasoning_effort. Set reasoning_effort to none with function tools, or use the Responses API.

Does CloudAxis still use 4o-mini?

Auto Mode is GPT-5.6 Luna on every plan. 4o-mini may still appear in some tool/fallback catalogs; product chat defaults to Luna.

Should I migrate if mini passes all my evals?

Not required. Stay on mini until you need Luna’s context, tools, or measured quality wins. Migration is optional, not mandatory.