Picking a model tier for an agent means deciding, for each step the agent takes, how much intelligence to pay for. The labs have made that decision sharper in the past two weeks. OpenAI shipped GPT-6.1 Terra on October 26, 2026, a fast tier at half the price of its siblings. Google announced Gemini 4 Carbon on October 28 as its strongest model for cybersecurity and long tasks, available only to a small group of defenders. One is cheap and available; the other is strong and locked up. This guide gives you a framework for routing agent work across tiers like these, and is honest about what we have not tested.
What did OpenAI ship with GPT-6.1 Terra?
The headline is price. GPT-6.1 Terra (API id gpt-6.1-terra) costs $1 per million input tokens and $5 per million output tokens, which OpenAI says is 50% below both GPT-5.6 Terra and GPT-6.1 Sol, the flagship that launched on September 29. It is available to Plus, Pro, Enterprise and Edu users in ChatGPT, in Codex and through the API.
The second claim is speed. OpenAI says Terra completes tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol, per BleepingComputer. For an agent making 30 tool calls per task, latency compounds.
The third is accuracy on a cheap setting. On an internal factual-error test at low reasoning effort, the error rate reportedly fell from 53.6% for GPT-5.6 Terra to 16.7% for GPT-6.1 Terra, according to Neowin. Low effort is where you would actually run a fast tier. Terra can also view a user’s screen and act across applications, so it is pitched at computer-use agents, not just chat.
OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra for the hardest coding work. Read that as the lab telling you where the tier boundary is.
What is Gemini 4 Carbon and can you use it?
Gemini 4 Carbon arrived two weeks after Gemini 4 Argon. Google calls it its strongest model for cybersecurity and long-running tasks. On the Artificial Analysis index, Carbon reportedly scores 59 against 58 for Claude Opus 5.5 and 57 for GPT-6 Astra, per Android Authority; most other scores come from Google’s own testing.
You probably cannot use it. Early access is limited to cyber defenders and Google has given no date for wider release or any pricing. For now, Carbon is a signal about where Google is putting its strongest model, not a tier you can route to.
For context, the same week brought DeepMind’s Genie 4, a real-time world model in research preview, and Mistral Large 4 reportedly shipped on October 29 per release trackers, unconfirmed by us.
The three tiers at a glance
We have not benchmarked any of these models ourselves. Prices and speed claims below come from the vendors.
| Tier | Example | Price (per 1M tokens) | Speed | When to use |
|---|---|---|---|---|
| Fast | GPT-6.1 Terra | $1 in / $5 out | OpenAI claims 3x faster than Sol and Astra | Tool-calling loops, computer use, routine steps |
| Flagship | GPT-6.1 Sol, GPT-6 Astra | Not public in our sources | Slower | Planning, verification, hardest coding |
| Specialist | Gemini 4 Carbon | Not public | Not public | Cybersecurity and long tasks, if you can get access |
How should you route agent steps across tiers?
An agent is not one model call. It is a sequence of steps with different difficulty, and the mistake most teams make is picking one model for all of them. Route by step type instead.
Planner and verifier on a top model
The step that decomposes the task and the step that checks the final answer carry the most risk. A wrong plan wastes every call that follows; a weak verifier lets errors through. These two steps run once or twice per task, so paying flagship prices for them is cheap in absolute terms. Put GPT-6.1 Sol or GPT-6 Astra here, or whichever top model your evals favor.
Tool-calling steps on the fast tier
The middle of an agent run is dozens of calls that each do something small: pick a tool, fill its arguments, read the result, decide the next move. This is where GPT-6.1 Terra belongs. The work is bounded, the cost of a single mistake is low if the verifier catches it, and the three-times speed claim turns a two-minute run into a tolerable one.
Extraction on the cheapest thing that passes
Pulling a date out of a document or classifying a ticket does not need a reasoning model. Use the cheapest tier that passes your eval set for that step, and revisit when a new cheap tier ships.
If you are new to this structure, our first agent tutorial shows a single-model agent you can split into tiers later.
Which metric should you track?
Cost per completed task, not cost per token. A cheap model that fails 30% of tasks and forces a retry on a flagship costs more than a mid-tier model that finishes the job the first time. Log every run with the model used at each step, total tokens, wall-clock time and whether the task passed verification. Divide total spend by passed tasks. That number is the only one worth optimizing.
For interactive agents, also track time to completion at the 90th percentile. A fast tier that saves 30 seconds per run is worth a few points of accuracy on steps the verifier covers.
Set a reasoning-effort policy
Most APIs let you set how hard the model thinks per request. Decide this per step, not per agent: low effort for tool selection and extraction, medium for synthesis, high only for planning and verification. OpenAI’s factual-error figure for Terra was measured at low effort, which is the setting that matters for the steps you would give it. Keep the policy in configuration, not scattered constants, so you can change it in one place when the next tier ships.
Keep an eval set and re-run it on every release
Tiers now ship every two to four weeks. Build a fixed set of 50 to 200 real tasks with known-good outputs and run it against each candidate model for each step type. When a new fast tier appears, run the tool-calling slice against it, compare cost per completed task, and switch or not based on numbers. Without this, you are choosing models from launch posts written by the vendor.
Watch for the cheap tier failing on long context
The failure mode readers report most often: a fast tier passes evals on short tasks, then starts inventing details once context grows past a few dozen tool results. It is not worse at any single step; it is worse at tracking what already happened. Guard against it by summarizing tool results before they enter context, capping how many steps run on the fast tier before a flagship checkpoint, and including long tasks in your eval set.
Our take
GPT-6.1 Terra is the more important release for working agent builders, because it is available and priced to be the default middle tier. If your agent runs every step on a flagship, moving the tool-calling loop to Terra is the obvious experiment, and your eval set will tell you within a day whether it holds up.
Gemini 4 Carbon is a statement of intent. The strongest models are increasingly gated for specific uses, cybersecurity first. Plan around what you can call today.
We will publish our own numbers once we have run our standard agent eval against Terra. Until then, treat every figure above as the vendor’s claim, build the routing structure, and let your data pick the models. See our tools page for current listings, or get updates in the newsletter.
Questions, answered
How much does GPT-6.1 Terra cost?
GPT-6.1 Terra is priced at $1 per million input tokens and $5 per million output tokens, which OpenAI says is 50% below GPT-5.6 Terra and GPT-6.1 Sol. The API model id is gpt-6.1-terra.
Should I use the cheapest model for my AI agent?
Not for every step. Use a top model for planning and verification, a fast tier like GPT-6.1 Terra for routine tool-calling steps, and the cheapest tier only for simple extraction. Then measure cost per completed task, not per token.
Is Gemini 4 Carbon available to developers?
Not generally. Google announced Gemini 4 Carbon on October 28, 2026 with early access limited to a small group of cyber defenders and no date for wider availability. Pricing has not been published.
