Hi, I'm Randika.
I'm an independent developer who handles model migrations, bug fixes and ongoing maintenance for apps built on any major AI provider: OpenAI, Claude, Gemini, Azure OpenAI, Vertex AI and Amazon Bedrock. When a provider retires a model, I find what breaks and move you to the replacement.
Nov 30, 2026
Anthropic retires claude-sonnet-4-5-20250929 (and claude-sonnet-4-5) on the Claude API. After that, requests to it fail. Recommended replacement: claude-sonnet-5-5.
Dec 1, 2026
On the Gemini API, requests to gemini-2.5-pro are routed to gemini-3.8-flash, and requests to gemini-2.5-flash to gemini-3.5-flash-lite, billed at those models' rates. Google says to review the breaking API changes first.
Jan 14, 2027
OpenAI shuts down GPT‑4o, GPT‑4o mini, GPT‑4.1 and GPT‑4.1 mini in the OpenAI API. After that, requests to them fail. Recommended replacement: gpt-6-luna.
After these dates, code that still names these models fails or quietly runs on a different model. I can help you migrate before then: I find every place your code depends on them, move it to the replacement, and check that it still works.
Every price is fixed and agreed before I start, and you pay after delivery.
Migration off GPT-4o and GPT-4.1
I find every place your code depends on a retiring model and send you a report.
1–2 days
The audit, plus I update the code and send you a pull request with a simple report.
2–4 days
Everything in Migration, plus I compare old vs new model outputs using your API key.
3–5 days
Other help
I fix bugs in apps that use the OpenAI, Claude or Gemini APIs, like failing tool calls, API errors and broken AI-built apps.
Price and time quoted per issue
I keep your AI features working as models get deprecated. I watch provider deprecation notices, scan your code with luna-check, and send fixes before anything breaks.
Plain-language results you can read, check and keep.
A report listing every file and line that depends on a retiring model, plus the hidden problems (tiktoken, temperature, tool calls) and what to change.
A pull request with the code changes and a MIGRATION.md that explains what changed, why, every file and line touched, and what was left alone on purpose.
Side-by-side results from the old and new model, run with your API key, so you can see the difference before you merge.
Changing the model name isn't enough. These are real problems I found while migrating open-source projects to gpt-6-luna, OpenAI's recommended replacement for gpt-4o-mini.
reasoning_effort: "none"gpt-6-luna is a reasoning model. OpenAI's model page says Chat Completions supports function calling only with reasoning_effort set to "none". Otherwise, tool calls need to move to the Responses API.
tiktoken crashestiktoken.encoding_for_model("gpt-6-luna") raises KeyError, in both tiktoken 0.7.0 and 0.14.0. Code that counts tokens breaks the moment you rename the model.
temperature has to goOpenAI's migration guide says to omit temperature, top_p and top_logprobs when moving to its newest reasoning models. It's often hidden in a shared client that always sends it.
# Initialize tokenizer- tokenizer = tiktoken.encoding_for_model("gpt-4o-mini")+ try:+ tokenizer = tiktoken.encoding_for_model(OPENAI_MODEL)+ except KeyError:+ tokenizer = tiktoken.get_encoding(FALLBACK_ENCODING) # "o200k_base"
o200k_base, the encoding gpt-4o-mini used. Chunk sizes stay the same.$ luna-check youtube_transcription_apiluna-check 0.2.0: scanned 2 file(s) in youtube_transcription_api app.py line 43 HIGH LC002 tiktoken.encoding_for_model() raises KeyError for gpt-6-luna Called with "gpt-4o-mini". tiktoken has no entry for "gpt-6-luna", so after the switch this raises KeyError (checked with tiktoken 0.7.0 and 0.14.0). Fix: Catch the error and fall back to get_encoding("o200k_base"), the encoding tiktoken uses for gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini. Source: https://github.com/openai/tiktoken/blob/main/tiktoken/model.py line 73 HIGH LC001 Model shuts down on Jan 14, 2027 "gpt-4o-mini" points to the dated snapshot(s) gpt-4o-mini-2024-07-18, which shut down on Jan 14, 2027. Replace it with "gpt-6-luna". Fix: Use "gpt-6-luna" (or read the model name from config). Source: https://developers.openai.com/api/docs/deprecations Found 2 high, 0 low in 1 file(s). Need help migrating? github.com/prasadrandika
Paste Python, JavaScript/TypeScript or Go code, or a config file (.env, YAML, JSON or TOML), and click Check. It runs the same 6 rules as luna-check 0.4.1, ported to JavaScript, and knows the retired and retiring models of OpenAI, Anthropic, the Gemini API, Vertex AI, Azure OpenAI and Amazon Bedrock.
Example: agents/books/llm.go from Nasiko as it was before my migration (Apache-2.0, license).
Your code never leaves your browser. The check runs on this page and nothing is uploaded.
These rules catch the problems I've seen most in real migrations. They're not a full review: a clean result means none of these patterns were found, not that the code is ready.
luna-check is free and open source (MIT). Run it on your own project to see what will break. It needs Python 3.9 or newer and Git.
pip install git+https://github.com/prasadrandika/luna-check@v0.4.1luna-check path/to/your/project
If it finds something and you'd like help fixing it, send me the output.
Real migration work on open-source projects. Statuses are as they stand today.
llm_prompt defaulted to gpt-4o-mini and the shared client always sent temperature. The maintainer confirmed it, fixed temperature upstream, and opened #382 with my commit. My commit is now part of ProjectDiscovery's dsl.claude-sonnet-4-5 to claude-sonnet-5-5 before Sonnet 4.5 retires on Nov 30, 2026, and checked that no sampling settings are sent to the new model.claude-sonnet-5-5, and fixed response parsing so a reply that starts with a thinking block is still read correctly.claude-sonnet-5-5, with tests that fail if a default model is ever marked deprecated.gpt-4o. I filed an issue with a ready branch.gpt-4o-mini, and tool calls need reasoning_effort: "none".gpt-4o-mini. Fixed, with a test.gpt-6-luna to the Agent Network's OpenAI model catalog before GPT-4o shuts down.temperature to the new model. The maintainer then made the model change directly, so my PR now only fixes outdated "GPT-4" mentions in the README.tiktoken. I added the fallback shown above and tested it end to end against a mock server.tiktoken.encoding_for_model() calls, and tool calls sent without reasoning_effort.Simple and low-risk for you.
Yes. I only need read access to your repository, or a zip of the code. I never store your API keys, and I'm happy to sign an NDA.
Read access to the repository (or a zip) is enough for an audit, a migration or a bug fix. For Migration + testing, I also run the old and new model with your API key so I can compare outputs.
I fix it at no extra cost.
PayPal or Payoneer, after delivery.
Sri Lanka (UTC+5:30). I reply within 24 hours.
Send me a short note about your stack. I prefer email or chat over calls, and I'll get back to you with a quick check and a fixed price.
Email meprasadrandika618 [at] gmail.com