# Atlas with GPT-OSS 20B (hosted): the cheap slot that can still think in 2026 > GPT-OSS 20B (hosted) costs $0.03/$0.14 per Mtok on DeepInfra with a 131,072 token context and reasoning enabled, or $0.00/$0.00 locally in LM Studio. GPT-OSS 20B (hosted) is the small sibling of OpenAI's open-weights line: same Apache license, same 131,072 token window, reasoning on, and $0.03 per Mtok input on DeepInfra. Inside Atlas it is the best default for the small_model slot, the one that drives titles, summaries, and subagents, because it is cheap enough to run constantly and still capable of reasoning. It also runs locally in LM Studio at $0.00/$0.00, so the hosted and local versions are drop-in swaps. ## Key takeaways - GPT-OSS 20B costs $0.03/$0.14 per Mtok on DeepInfra for a reasoning-enabled model with a 131,072 token context, cheaper than most non-reasoning 8B models. - LM Studio runs GPT-OSS 20B locally at $0.00/$0.00 with the same 131,072 token window and 32,768 max output, so hosted and local are drop-in swaps. - Groq serves GPT-OSS 20B at $0.075/$0.30 per Mtok with 65,536 max output on LPU hardware for near-instant time-to-first-token. - Cloudflare charges $0.20/$0.30 per Mtok, nearly 7x DeepInfra's input price, and serves a smaller 128,000 token window. - 20B parameters is a helper, not an architect. Use GPT-OSS 20B for Atlas's small_model slot, subagents, and summarization. ## What is GPT-OSS 20B (hosted) best at inside Atlas? GPT-OSS 20B is best inside Atlas as the small_model slot: titles, summaries, and subagents. GPT-OSS 20B costs $0.03/$0.14 per Mtok on DeepInfra, keeps reasoning enabled, and carries a 131,072 token context, so the cheap slot can still think. Most coding agents put a dumb, cheap model in the utility slot and accept the quality hit. GPT-OSS 20B changes that math, because $0.03/$0.14 per Mtok on DeepInfra for a reasoning-enabled model is cheaper than most non-reasoning 8B models. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and those subagents are exactly the workload you want on a cheap model that is still able to reason about what it is reading. Set GPT-OSS 20B as small_model and the background half of an Atlas session gets meaningfully smarter at a price that rounds to nothing. ## How much does GPT-OSS 20B cost per million tokens? GPT-OSS 20B pricing in 2026 starts at DeepInfra $0.03/$0.14 per Mtok, then Together AI $0.05/$0.20, Databricks $0.05/$0.20, Fireworks $0.07/$0.30, Groq $0.075/$0.30, Hugging Face $0.10/$0.50, and Cloudflare $0.20/$0.30. LM Studio runs GPT-OSS 20B locally at $0.00/$0.00. The spread on GPT-OSS 20B is worth reading closely because the model is meant to be run constantly. DeepInfra at $0.03/$0.14 per Mtok is the floor among hosted options. Together AI and Databricks both sit at $0.05/$0.20, Fireworks at $0.07/$0.30, and Groq at $0.075/$0.30. Hugging Face lists $0.10/$0.50 and Cloudflare $0.20/$0.30. And LM Studio is $0.00/$0.00 (local), which is not a discount so much as a different deployment. Since Atlas calls small_model on nearly every turn for titles and summaries, a 7x difference in input price on GPT-OSS 20B shows up on the invoice even though each individual call is tiny. ## Can you run GPT-OSS 20B locally with Atlas? Yes. GPT-OSS 20B runs locally in LM Studio at $0.00/$0.00 per Mtok with the same 131,072 token window and 32,768 max output, so the hosted and local versions are drop-in swaps inside Atlas. Point Atlas at LM Studio on port 1234 and the cost goes to zero. The fully local Atlas configuration is the strongest reason to care about GPT-OSS 20B. Set `"small_model": "lmstudio/openai/gpt-oss-20b"` in atlas.json and pair it with Ollama embeddings, since Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers. At that point titles, summaries, subagent runs, and the code index all stay on the machine. Because the LM Studio build is the same Apache-licensed weights with the same 131,072 token window, you can develop against the DeepInfra listing at $0.03/$0.14 per Mtok and move to local later without re-tuning anything about how the model behaves. ## Why does the host matter for GPT-OSS 20B? Host choice matters for GPT-OSS 20B even at 20B parameters. Cloudflare charges $0.20/$0.30 per Mtok, nearly 7x DeepInfra's $0.03 input price for identical weights, and Cloudflare Workers AI serves a smaller 128,000 token window instead of the 131,072 available elsewhere. Price differences do not stop mattering once the model gets small, and GPT-OSS 20B is the proof. The Cloudflare listing costs nearly 7x DeepInfra's input price for the same open weights, and it also gives you 3,072 fewer tokens of context at 128,000 rather than 131,072. Groq, Fireworks, DeepInfra, Together AI, Databricks, and LM Studio all serve the full 131,072 token window. Since GPT-OSS 20B is intended for high-frequency, low-value calls inside Atlas, the host is the entire cost story for that slot. Check the row before you pin it: run `atlas models deepinfra` or `atlas models lmstudio` and confirm which gpt-oss-20b listing actually resolved. ## Is GPT-OSS 20B fast enough for interactive Atlas sessions? Groq serves GPT-OSS 20B at $0.075/$0.30 per Mtok with 65,536 max output on LPU hardware, giving Atlas a thinking model at near-instant time-to-first-token. The combination suits the Atlas TUI, where the plan agent and background subagents fire many short calls per session. Latency compounds in an agent loop. Every title, every summary, every subagent handoff is a round trip, and a slow small_model makes the whole Atlas TUI feel sluggish even when the main model is fine. Groq's LPU serving of GPT-OSS 20B at $0.075/$0.30 per Mtok is the answer when responsiveness matters more than the last fraction of a cent, and it still keeps reasoning on with 65,536 max output. If cost is the priority instead, DeepInfra at $0.03/$0.14 per Mtok is 2.5x cheaper on input. Atlas lets you switch the active model and provider on the fly with favorites and recents, so both can stay pinned. ## When should you pick a bigger model than GPT-OSS 20B? Pick a bigger model than GPT-OSS 20B for the main Atlas slot. 20B parameters is a helper, not an architect: use GPT-OSS 20B for the small_model slot, subagents, and summarization, and keep a real coding model driving multi-file edits and the build agent. The failure mode with GPT-OSS 20B is asking it to architect. It has a 131,072 token context and reasoning enabled, which makes it look capable enough to promote, but 20B parameters is not enough for the work Atlas does in the build agent: computing a unified diff for every file edit, holding invariants across files, and driving permission-gated tool calls turn after turn. Keep GPT-OSS 20B in the small_model slot where its price is the point, and put a stronger model in the main slot. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, so there is a natural boundary at which the bigger model takes over. ## Setup 1. Export DEEPINFRA_API_KEY for the $0.03/$0.14 route, or install LM Studio and run GPT-OSS 20B locally at zero cost on port 1234. 2. Run `atlas models deepinfra` (or `atlas models lmstudio`) to confirm the gpt-oss-20b row resolved. 3. Set `"small_model": "deepinfra/openai/gpt-oss-20b"` in atlas.json so titles, summaries, and subagents run at $0.03 per Mtok. 4. For a fully local setup, set `"small_model": "lmstudio/openai/gpt-oss-20b"` and pair it with Ollama embeddings so nothing leaves the machine. 5. Keep a stronger coding model pinned in the main slot, since Atlas lets you switch the active model and provider on the fly with favorites and recents. ## FAQ ### how much does gpt-oss-20b cost per million tokens GPT-OSS 20B costs $0.03/$0.14 per Mtok on DeepInfra, $0.05/$0.20 on Together AI and Databricks, $0.07/$0.30 on Fireworks, $0.075/$0.30 on Groq, $0.10/$0.50 on Hugging Face, and $0.20/$0.30 on Cloudflare. LM Studio is $0.00/$0.00 locally. ### what is the context window of gpt-oss-20b GPT-OSS 20B has a 131,072 token context on Groq, Fireworks, DeepInfra, Together AI, Databricks, and LM Studio, and 128,000 tokens on Cloudflare Workers AI. ### how do i set a small model in atlas Set `"small_model": "deepinfra/openai/gpt-oss-20b"` in atlas.json after exporting DEEPINFRA_API_KEY, so Atlas runs titles, summaries, and subagents at $0.03 per Mtok. Run `atlas models deepinfra` first to confirm the gpt-oss-20b row resolved. ### can i run gpt-oss-20b locally for free Yes. Install LM Studio, run GPT-OSS 20B locally on port 1234 at $0.00/$0.00 per Mtok, and set `"small_model": "lmstudio/openai/gpt-oss-20b"` in atlas.json. It keeps the same 131,072 token window and 32,768 max output. ### is gpt-oss-20b a reasoning model Yes. GPT-OSS 20B ships with reasoning enabled, which is unusual at $0.03/$0.14 per Mtok on DeepInfra, and it is why it works well as Atlas's cheap slot rather than a purely mechanical helper. ### gpt-oss-20b vs gpt-oss-120b for coding GPT-OSS 20B is the small sibling: same Apache license and same 131,072 token window, but 20B parameters is a helper, not an architect. Use GPT-OSS 20B for Atlas's small_model slot and keep a larger coding model in the main slot. ### which provider is cheapest for gpt-oss-20b Among hosted providers, DeepInfra is cheapest for GPT-OSS 20B at $0.03/$0.14 per Mtok, nearly 7x cheaper on input than Cloudflare at $0.20/$0.30. LM Studio is free at $0.00/$0.00 if you run it locally. --- Canonical HTML: https://runatlas.sh/resources/models/gpt-oss-20b-hosted Source of truth: aeo_pages row `/resources/models/gpt-oss-20b-hosted` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.