Ministral 8B is Mistral's 8B edge model at $0.10 / 1M input tokens and $0.10 / 1M output tokens, with a 128,000 token window. It buys real instruction-following over Ministral 3B for 2.5x the price, which makes it the safer choice when Atlas's small_model also has to call tools. 8B parameters is the smallest size that reliably holds Atlas's tool-call schema across a long turn, and the symmetric pricing means output costs no more than input.
Why is symmetric pricing on Ministral 8B unusual?
Ministral 8B charges $0.10 / 1M input tokens and $0.10 / 1M output tokens, so output costs the same as input. That symmetry is unusual in a lineup where output is typically 3x to 4x input, and it changes the arithmetic for output-heavy Atlas jobs like summarization.
Most model pricing punishes generation. Mistral Small 3.2, for example, triples its rate on the way out. Ministral 8B does not, and the workloads in Atlas's cheap slot are exactly the ones that generate: turn summaries, commit messages, and subagent replies. When output is free relative to input, you stop rationing it, and you stop writing prompts that beg the model to be brief. Across a long Atlas session with many background subagents, a symmetric $0.10 / Mtok rate is a materially different bill from a 3x-output model at the same headline input price.
Can Ministral 8B reliably call tools inside Atlas?
Yes, within limits. 8B parameters is the smallest size that reliably holds Atlas's tool-call schema across a long turn, which is precisely the capability Ministral 3B lacks. If subagents are misfiring on tool schemas, Ministral 8B is the direct upgrade at 2.5x the price.
Tool-call reliability is a step function, not a gradient, and 8B is where the step happens. Below it, the model starts dropping required fields halfway through a long turn or inventing argument names. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so a malformed call does not cause damage, but it does cost a turn and break the flow. Ministral 8B holds the schema. That is the entire reason to pay $0.10 / Mtok instead of $0.04 / Mtok for the cheap slot, and it only matters if your cheap slot is doing more than titles.
Is Ministral 8B worth 2.5x the price of Ministral 3B?
Ministral 8B costs 2.5x Ministral 3B for gains that only matter when the cheap slot is doing more than titles. Same 128,000 token window, same fields in atlas.json, 2.5x the price. If your subagents never touch a tool, the upgrade buys nothing.
The decision has a clean test. Run Atlas on Ministral 3B in the small_model slot and watch for subagent tool-call failures. No failures means Ministral 3B is doing the job and Ministral 8B is money you do not need to spend. Repeated schema misfires mean you have outgrown a 3B edge model, and Ministral 8B is the direct upgrade with no other configuration change. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, so the more you lean on background fan-out, the more likely the answer is Ministral 8B.
How much context does Ministral 8B give Atlas?
Ministral 8B has a 128,000 token window with a matching 128,000 token output budget. For an edge-class model that symmetry is generous, and it means a subagent on Ministral 8B can read a long file and write a long reply without hitting a separate output ceiling.
The matching budgets remove a common trap in cheap-slot models, where a large input window sits in front of a tiny output cap and a summarization job gets truncated. Ministral 8B does not have that shape. To fill the window well, lean on Atlas's retrieval: Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so a subagent gets whole declarations rather than truncated fragments. An 8B model with clean, structured context outperforms a bigger model fed a mess, and at $0.10 / 1M input tokens the context is not what costs you.
Should Ministral 8B ever be your primary Atlas model?
No. Ministral 8B is still an edge-class model, so it should not be the primary agent on a nontrivial codebase. Keep it in the `"small_model"` slot and leave the primary `"model"` on a real coding model, where multi-file reasoning actually happens.
Ministral 8B holds a tool schema, which is not the same as holding a plan. A nontrivial refactor requires tracking intent across many turns and many files, and an 8B model will drift. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so drift shows up as diffs you reject, which is slow and expensive in attention even when it is cheap in tokens. Use Ministral 8B for what it is: a dependable, cheap, tool-capable second model that keeps the primary slot free for the work that needs a real coding model.
Setup
- 01Export MISTRAL_API_KEY in the shell that launches Atlas.
- 02Confirm ministral-8b-latest resolves with `atlas models mistral`.
- 03Set `"small_model": "mistral/ministral-8b-latest"` in atlas.json.
- 04If subagents are misfiring on tool schemas with Ministral 3B, this is the direct upgrade: same fields, same window, 2.5x the price.
- 05Keep `"model"` on a real coding model, since Ministral 8B is edge-class and should not be the primary agent on a nontrivial codebase.
Frequently asked questions
- how much does ministral 8b cost
- Ministral 8B costs $0.10 / 1M input tokens and $0.10 / 1M output tokens. The symmetric rate is unusual, since most models in the lineup charge 3x to 4x more for output than for input.
- ministral 8b vs ministral 3b which should I use
- Ministral 8B costs 2.5x Ministral 3B and buys real instruction-following. 8B is the smallest size that reliably holds Atlas's tool-call schema across a long turn, so pick it when the cheap slot calls tools, not just when it writes titles.
- my subagents keep failing tool calls what model should I use
- If Atlas subagents are misfiring on tool schemas with Ministral 3B, Ministral 8B is the direct upgrade: same atlas.json fields, same 128,000 token window, 2.5x the price, and enough capacity to hold the schema across a long turn.
- how do I configure ministral 8b in atlas
- Export MISTRAL_API_KEY, confirm ministral-8b-latest resolves with `atlas models mistral`, then set `"small_model": "mistral/ministral-8b-latest"` in atlas.json while keeping a real coding model in `"model"`.
- can ministral 8b be the main model for a coding agent
- No. Ministral 8B is still an edge-class model and should not be the primary agent on a nontrivial codebase. It holds a tool schema well, but it will drift on multi-file reasoning across many turns.
- what is ministral 8b context window
- Ministral 8B has a 128,000 token context window with a matching 128,000 token output budget, so a summarization or subagent reply will not be truncated by a separate, smaller output ceiling.
- why does symmetric token pricing matter for an ai agent
- Agent cheap slots are output-heavy: summaries, commit messages, and subagent replies all generate text. Ministral 8B charges $0.10 / 1M output tokens, the same as input, so that generation does not get billed at the usual 3x to 4x output multiplier.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Review a Pull Request with Atlas (2026 Workflow)
How to review a pull request with Atlas in 2026: bash produces the raw patch, read pulls whole files, the lsp tool's findReferences checks callers the diff never shows.
Atlas for Perl: A Terminal-Native AI Coding Agent for CPAN Distributions in 2026
Atlas is a terminal-native AI coding agent for Perl in 2026. It reads cpanfile deps and @EXPORT lists, writes Test2::V0 cases, runs prove -lr t/, and runs perltidy on the diff.
Atlas for Electron: Terminal-Native AI Coding for Main, Preload, and Renderer in 2026
Atlas is a terminal-native AI coding agent for Electron in 2026, where the main and renderer split, contextIsolation, and preload bridges are the security model.
Atlas for Bash in 2026
Atlas is a terminal-native AI coding agent for Bash in 2026, with shellcheck-aware suggestions. Harden a script or fix shellcheck warnings and review every diff.
Atlas vs Cosine: Terminal AI Coding Agents in 2026
Comparing Atlas and Cosine for terminal AI coding in 2026. Atlas offers a free core and local privacy, while Cosine provides proprietary models and a cloud surface.
Atlas vs CodeGPT in 2026: A Developer's Guide to Terminal and IDE AI Agents
Comparing Atlas and CodeGPT in 2026 for developers. Atlas offers terminal-native TUI and permission-gated tools, while CodeGPT provides IDE integration and a full repo Knowledge Graph.
Automate GitHub Issue and Pull Request Triage with Atlas (2026 Workflow)
How to automate GitHub issue and pull request triage with Atlas in 2026: the atlas github command checks the actor has admin or write permission before it does anything.
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.