Baseten is a dedicated-deployment host, and it shows in the shape of the catalog: every model has max output equal to its full context, 262,000 in and 262,000 out on Kimi K2.7 Code, because you are not sharing a throttled pool with everyone else. Inside Atlas, the terminal-native coding agent, Baseten is the pick when output truncation is your real risk. Pricing runs GPT OSS 120B $0.10/$0.50 per Mtok, Kimi K2.5 $0.60/$3.00, Kimi K2.7 Code $0.95/$4.00, Nemotron Ultra $0.60/$2.40, GLM 4.7 $0.60/$2.20, and GLM 5.2 $1.40/$4.40.
What is Baseten best at inside Atlas?
Baseten is best inside Atlas for long agentic edit sessions where output truncation is the real risk. Max output equals the full context on most Baseten models, so a 262,000 token Kimi K2.7 Code call can theoretically emit 262,000 tokens, an output ceiling no other 2026 host matches.
That property is a direct consequence of Baseten being a dedicated-deployment host rather than a shared pool. Compare the same weights elsewhere: DeepInfra caps Kimi K2.7 Code output at 262,144, and GLM-5.1 at just 16,384. For Atlas the output number is the one that bites. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so a truncated emission is a truncated patch, and a truncated patch is a failed turn. Baseten removes that failure mode structurally on Kimi K2.6, Kimi K2.5, Kimi K2.7 Code, and DeepSeek V4 Pro, all at 262,000 tokens with 262,000 max output.
How much does Baseten cost per million tokens?
Baseten prices GPT OSS 120B at $0.10/$0.50 per Mtok, Kimi K2.5 at $0.60/$3.00, Kimi K2.7 Code at $0.95/$4.00, Nemotron Ultra at $0.60/$2.40, GLM 4.7 at $0.60/$2.20, and GLM 5.2 at $1.40/$4.40. GPT OSS 120B at $0.10/$0.50 is roughly a third of the Groq price.
GPT OSS 120B on Baseten at $0.10/$0.50 per Mtok lands well under Groq ($0.15/$0.60) and well under Cerebras ($0.35/$0.75), which makes it a strong `small_model` for Atlas. Nemotron Ultra is the other pricing story: at $0.60/$2.40 per Mtok it undercuts the Together AI listing of the same 550B A55B model ($0.60/$3.60) on output by 33 percent. Output pricing is where an Atlas agent actually spends, because diffs, plans, and subagent summaries are all output tokens. A 33 percent cut on the output side of a 550B model is a meaningful line item on a long build session.
What are the limits of the Baseten catalog for Atlas?
Baseten carries only 13 models in the Atlas registry, all open weights. There is no Claude, no GPT-5, and no Gemini on Baseten, so you cannot use Baseten as your only provider if any part of your workflow depends on a frontier closed model in 2026.
Run `atlas models baseten` and you will see the whole catalog in one screen. That is a feature if you want Kimi, GLM, Nemotron, DeepSeek, and GPT OSS at dedicated-capacity output limits. It is a hard stop if you need frontier closed weights. Atlas lets you switch the active model and provider on the fly with favorites and recents, so the usual answer is to run Baseten as one provider among several: Kimi K2.7 Code from Baseten for the long build sessions where 262,000 max output matters, and a frontier provider on the favorites list for the turns that need it.
What is the downside of Baseten's dedicated capacity?
Baseten's generous output limits are a property of dedicated capacity, and dedicated capacity means cold starts and per-deployment configuration are things you now have to think about. A shared-pool provider hides that from you. Baseten does not, which is the price of 262,000 max output.
If your Atlas usage is spiky, a short session in the morning and another after lunch, cold starts are a real cost that does not appear on any per-Mtok price list. If your usage is a sustained multi-hour build session, the dedicated capacity is exactly what you want and the cold start is amortized away in the first minute. Baseten is therefore a poor fit for a quick one-off question and a strong fit for the long agentic session Atlas is built around, where Atlas fans out work to subagents that can run in the foreground or in parallel background sessions.
How do I configure Atlas to run on Baseten?
Configure Atlas for Baseten by exporting BASETEN_API_KEY from the Baseten dashboard. Atlas hits https://inference.baseten.co/v1 through @ai-sdk/openai-compatible, so no custom SDK is needed. Run `atlas models baseten` to list the 13 resolved models before pinning one.
For the coding slot, pin `"model": "baseten/moonshotai/Kimi-K2.7-Code"` in atlas.json. That is the row where Baseten's shape pays off: long agentic edit sessions where output truncation is the real risk, backed by 262,000 tokens of context and 262,000 max output. For the cheap slot, set `"small_model": "baseten/openai/gpt-oss-120b"` at $0.10/$0.50 per Mtok, roughly a third of the Groq price. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, which pairs well with a model that can emit a full patch without truncating.
When should I pick a different provider than Baseten?
Pick a different provider than Baseten when you need Claude, GPT-5, or Gemini, because Baseten carries only 13 models in the Atlas registry and all of them are open weights. Also pick elsewhere if your Atlas usage is spiky, since Baseten's dedicated capacity means cold starts.
Baseten's argument is narrow and strong: nowhere else in the 2026 Atlas registry does max output equal the full context on a 262,000 token model. If that is not your bottleneck, the argument mostly evaporates and you are choosing on price alone, where GPT OSS 120B at $0.10/$0.50 per Mtok and Nemotron Ultra at $0.60/$2.40 are still competitive but not unique. Ask yourself whether your Atlas sessions are failing because a model truncated a large multi-file diff. If yes, Baseten is the answer. If no, pick on price, catalog breadth, or latency instead.
Setup
- 01Export BASETEN_API_KEY from the Baseten dashboard. Atlas hits https://inference.baseten.co/v1 through @ai-sdk/openai-compatible.
- 02Run `atlas models baseten` to list the 13 resolved models, all open weights.
- 03Pin `"model": "baseten/moonshotai/Kimi-K2.7-Code"` in atlas.json for long agentic edit sessions where output truncation is the real risk.
- 04Set `"small_model": "baseten/openai/gpt-oss-120b"` at $0.10/$0.50 per Mtok for the cheap slot.
- 05Keep a second provider configured as an Atlas favorite, since Baseten carries no Claude, GPT-5, or Gemini.
Frequently asked questions
- how to use baseten with atlas
- Export BASETEN_API_KEY from the Baseten dashboard, run `atlas models baseten` to list the 13 resolved models, then pin `"model": "baseten/moonshotai/Kimi-K2.7-Code"` in atlas.json. Atlas hits https://inference.baseten.co/v1 through @ai-sdk/openai-compatible.
- how much does baseten cost per million tokens
- GPT OSS 120B is $0.10/$0.50 per Mtok, Kimi K2.5 is $0.60/$3.00, Kimi K2.7 Code is $0.95/$4.00, Nemotron Ultra is $0.60/$2.40, GLM 4.7 is $0.60/$2.20, and GLM 5.2 is $1.40/$4.40.
- which provider has the highest output token limit for coding agents
- Baseten. Max output equals the full context on most Baseten models, so a 262,000 token Kimi K2.7 Code call can theoretically emit 262,000 tokens. DeepInfra caps the same model's output at 262,144 and GLM-5.1 at just 16,384.
- does baseten have claude or gpt-5
- No. Baseten carries only 13 models in the Atlas registry, all open weights. There is no Claude, GPT-5, or Gemini, so Baseten cannot be your only provider if you need a frontier closed model.
- why does baseten have cold starts
- Baseten is a dedicated-deployment host, and the generous output limits are a property of that dedicated capacity. The tradeoff is that cold starts and per-deployment configuration are things you now have to think about, unlike a shared-pool provider.
- what is the best small model on baseten for an agent
- GPT OSS 120B at $0.10/$0.50 per Mtok. Set `"small_model": "baseten/openai/gpt-oss-120b"` in atlas.json. That price is roughly a third of the Groq listing ($0.15/$0.60) and well under Cerebras ($0.35/$0.75).
- is baseten cheaper than together ai for nemotron ultra
- On output, yes. Nemotron Ultra is $0.60/$2.40 per Mtok on Baseten versus the Together AI listing of the same 550B A55B model at $0.60/$3.60, an output saving of 33 percent. Input is the same $0.60.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Zed: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with Zed, the GPU-accelerated collaborative editor, for developers in 2026. Evaluate their AI autonomy, privacy, and workflow.
Atlas for SQL in 2026
Atlas is a terminal-native AI coding agent for SQL in 2026. Run it in a repo with your migrations or .sql files, optimize a query, and review the diff before applying.
Atlas for Polars: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Polars. Build LazyFrame chains, push scan_csv predicates into the reader, and read explain() plans in 2026.
Extract a Shared Helper from Duplicated Code with Atlas (2026 Workflow)
How to extract a shared helper from duplicated code with Atlas in 2026: codebase_search finds the copies by meaning, write creates the module, apply_patch swaps each call.
Atlas vs Claude Code: Terminal AI Coding Agents in 2026
Compare Atlas and Claude Code in 2026 for terminal AI coding. Atlas offers free core, BYO model keys, and robust diffing, while Claude Code provides deep Claude tuning and strong multi-step execution.
Atlas vs Roo Code: Terminal AI Coding Agents in 2026
Comparing Atlas and Roo Code in 2026. Atlas offers terminal-native TUI, permission-gated tools, and diff review. Roo Code, a VS Code extension, shut down May 15, 2026.
Migrate a Deprecated API Across Every Callsite with Atlas (2026 Workflow)
How to migrate a deprecated API across every callsite with Atlas in 2026: the lsp tool's findReferences enumerates callers, todowrite tracks them, apply_patch migrates each one.
Atlas for Actix Web in 2026
Atlas is a terminal-native AI coding agent for Actix Web in 2026. It reads extractors and app_data, then runs cargo test and cargo clippy behind a prompt.