Which AI models does Autoesta use?
Autoesta uses current models from OpenAI, Anthropic, Google and other providers, chosen per client for answer quality, speed, cost and data handling. Every new build uses a model with no announced retirement date.
There is no single "Autoesta model". The model is a line in each client's written plan, picked after we test the job on real cases, and it runs on the client's own provider account. The full service, with running costs per 1,000 tasks and the frameworks we build on, is on our custom AI agent development page. This page answers the narrower question buyers ask first: which models, how we choose, and what happens when a provider retires one.
We do not publish which model any client uses. That choice belongs to the client and depends on their job, their data rules and their budget.
Which AI models are current at OpenAI, Anthropic and Google right now?
On October 30, 2026, OpenAI lists GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna as its flagship models. Anthropic lists Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. Google lists Gemini 4 Argon, Gemini 3.8 Flash and Gemini 3.5 Flash-Lite as stable.
We read each provider's own models page and deprecations page on October 30, 2026. The table lists text models we would consider for a new agent or automation step, and what each provider's page says about retirement. Speech, image and video models are left out.
| Provider | Current text models on its models page | What its own page says about retirement |
|---|---|---|
| OpenAI | GPT-6 Astra (most capable), GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna (most efficient, high volume) | None of the four has a shutdown date on the deprecations page. Generally available models get "at least 6 months" notice. |
| Anthropic | Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5 | Retirement "not sooner than" September 1, 2027 (Fable 5.1) to October 7, 2027 (Haiku 5.5) |
| Gemini 4 Argon (flagship reasoning), Gemini 3.8 Flash, Gemini 3.5 Flash-Lite, all marked Stable | Gemini 3.8 Flash and 3.5 Flash-Lite show "No shutdown date announced"; Gemini 4 Argon is not in the deprecation tables |
Two notes from the same pages. Anthropic's models overview says "if you're unsure which model to use, start with Claude Opus 5.5 for most workloads," and describes Haiku 5.5 as built "for high-volume, latency-sensitive tasks such as classification, extraction, and routing." Google's page marks Gemini 3.1 Pro as a preview, and we do not put preview models into a client's live system, because previews can be withdrawn at short notice.
Other providers are in scope too when a client needs them, for example because a company has already approved one provider for its data. The same rule applies: current model, no announced retirement date.
What does "no announced retirement date" rule out today?
On October 30, 2026 it rules out at least 15 text models with published shutdown dates, including GPT-4o, GPT-4.1, the GPT-5 family, o3, Claude Sonnet 4.5 and Gemini 3.1 Flash-Lite, even though most of them still work today.
This is the practical effect of the rule. A model that still answers today but has a shutdown date would force a migration, a retest and possibly a price change within months of launch. The dates below come from our LLM API pricing and model retirement tracker, built from the same providers' pages.
| Shutdown date | Models we would not start a new build on | Provider's suggested replacement |
|---|---|---|
| November 30, 2026 | Claude Sonnet 4.5 | Claude Sonnet 5.5 |
| December 11, 2026 | GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 pro, o3, o3-pro | GPT-5.6 Sol, Terra or Luna |
| January 14, 2027 | GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini | GPT-6 Luna |
| April 1, 2027 | GPT-5.1, GPT-5.4 nano, GPT-5.3 Codex | GPT-6 Sol or GPT-6 Luna |
| May 7, 2027 (earliest) | Gemini 3.1 Flash-Lite | Gemini 3.5 Flash-Lite |
GPT-4o and GPT-4.1 matter most because many older automations still call them. We covered that notice in OpenAI retiring GPT-4o and GPT-4.1 on January 14, 2027. A price on a provider's pricing page is not proof a model is safe to build on; the deprecations page is the one to check.
How does Autoesta choose the model for a client?
We write down what a correct result looks like, build a test set from the client's real past cases, run it on a small current model first, and move to a larger one only for the steps that fail. Data rules can narrow the list before any test.
| Client requirement | What we check | How it changes the choice |
|---|---|---|
| Answer quality | Pass rate on the test set, including tool calls in the run log | A step that fails on a small model moves up a tier; the others stay small |
| Speed | Response time per step, which matters for chat and voice | Customer-facing replies favour the faster models in a family |
| Cost | Tokens per task from the test runs, times the provider's list price | The cheapest model that passes wins; the money page shows a twentyfold spread per 1,000 tasks |
| Data handling | Which providers the client allows, region, and the provider's data terms | Can limit the choice to one provider before testing starts |
The test set is the important part. A model that writes a good reply can still skip a tool or fill a field wrongly, so we grade the outcome and the run log, not just the text. Our guide on how to test an AI agent before launch sets out the method, and why AI agents are not deterministic explains why we run each case more than once.
How does an AI build with Autoesta run, step by step?
A free strategy call, a free first audit, a written plan with a fixed price that names the model, the build on your accounts, testing on real cases, handover, then free fixes to anything we built for 6 months.
- Free strategy call. You show us the job and the systems involved.
- Free audit. We look at the data, any existing AI steps and the model IDs they call.
- Written plan and fixed price. The tools, permissions, model and test set, written down before work starts.
- Build on your accounts. Your provider keys, your automation tool or repository, your CRM.
- Test. The test set runs on the chosen model, and on one alternative where cost or quality is close.
- Handover. Notes on which model each step uses and why.
- Six months of free fixes. Optional monthly support after that.
How much does an AI build with Autoesta cost?
At the time of writing, the strategy call and first audit are free, a small build of two or three automations is about $1,000, full setups run $2,000 to $5,000, and larger systems get a fixed price after the audit.
| What you need | Autoesta price at the time of writing |
|---|---|
| Strategy call and first audit | Free |
| Cleanup of an existing account | From $200 |
| Small build: two or three automations | About $1,000 |
| Full setup | $2,000 to $5,000 |
| Larger systems | Fixed price after the audit, shown before work starts |
| Fixes to anything we built | Free for 6 months |
| Monthly support after that | Optional, from $200 a month |
Model usage is separate. It is billed by the provider to your own account at its list price, so you see every token on your own dashboard. If you want help choosing where AI belongs before any build, that is paid AI consulting, quoted per business; our page on what Autoesta charges for AI consulting has the detail.
What happens when a provider retires a model Autoesta used?
We switch before the shutdown date, not on it. We rerun the client's test set on the provider's replacement and one cheaper alternative, then move to the one that passes. Within 6 months of launch that is a free fix.
Providers give different notice. OpenAI's deprecations page promises at least 6 months for generally available models, yet its October 14, 2026 notice for GPT-4o and GPT-4.1 gave 92 days. Anthropic's current models carry "not sooner than" dates in 2027. Google's shutdown dates are "the earliest possible dates on which a model might be retired." After the free 6 months, optional monthly support from $200 covers watching these notices and moving the model; see what support looks like after an automation build.
When do you not need Autoesta to pick an AI model?
You do not need us when the tool already picks the model for you, when the job is one prompt you can check by eye, or when your team already tests models on real cases before switching.
- Built-in AI features. Many helpdesks and CRMs run their own AI on a model they choose and update. You manage the feature, not the model.
- One low-risk prompt. Drafting a summary that a person reads before it goes anywhere rarely needs a test set. Pick a current model from the table above and check the provider's deprecations page once a quarter.
- An in-house team that already does this. If someone owns your model list, test cases and deprecation notices, a free audit may be all you want from us.
When the agent writes to your CRM, answers customers, or calls internal tools through an MCP server Autoesta builds, the model choice is worth testing properly.
More questions about Autoesta and AI models
Does Autoesta use the same AI model for every client?
No. The model is chosen per client and often per step, based on answer quality on that client's test set, response speed, cost per task and the client's data rules. One system can use a small, fast model for sorting and a larger one for a step that needs more reasoning. The choice is written into the plan, so you always know which model each step calls.
Can we tell Autoesta which AI provider to use?
Yes. If your company has already approved a provider, for example because of a data agreement, we build on that provider and pick the best fit among its current models. We still apply the same rule: no model with an announced retirement date. If none of that provider's current models passes the test set for a step, we tell you before the build and explain the options.
Can we use our own OpenAI, Anthropic or Google API key?
Yes, and we recommend it. The system runs on keys in your own provider account, so usage is billed to you at the provider's list price, you see the usage dashboard yourself, and you can rotate or remove the key at any time. Our build and support prices are separate from token costs. Ownership of keys, data and prompts is covered in who owns automations and data.
Will Autoesta move our old automations off a retiring model?
Yes. The free audit lists every model ID your agents and AI steps call and checks each against the providers' deprecation pages. We then rerun your real cases on the replacement and one cheaper alternative and switch before the shutdown date. At the time of writing, cleanup of an existing account starts from $200, and larger systems get a fixed price after the audit.
Does a newer model always cost more?
No. Some replacements are far cheaper: on October 30, 2026 prices, GPT-6 Luna lists at $0.10 per million input tokens against $2.50 for GPT-4o. Others cost more, so we price the switch on your real token use before moving. Our model retirement research shows the before and after cost for each retiring model. More answers about how we work are in the Autoesta FAQ.
Sources and further reading
- OpenAI, Models (flagship models), checked October 30, 2026
- OpenAI, Deprecations (shutdown dates, notice policy), checked October 30, 2026
- Anthropic, Models overview (current models, retirement dates), checked October 30, 2026
- Google, Gemini API models, checked October 30, 2026
- Google, Gemini API deprecations, checked October 30, 2026
