Research
Cursor Model Roadmap Tracker
At Compile on June 16, 2026, Cursor said it was training a significantly larger model from scratch with SpaceX using 10x more total compute. That is a dated roadmap announcement, not a current picker claim. Use the current model guide for today's Grok 4.6, Grok 4.5, Composer 2.5, Claude 5, Gemini 3.7 Flash, Gemini 3.1 Pro and GPT-5.6 choices.
On this page
What changed in Cursor's larger model work?
At Compile on June 16, 2026, Cursor said it was training a significantly larger model from scratch with SpaceX using 10x more total compute. That is a dated roadmap announcement, not a current picker claim. Use the current model guide for today's Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition., Grok 4.5, Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition., Claude 5, Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition., Gemini 3.1 ProGoogle's frontier model in Cursor for work that combines images and code, especially UI implementation and visual code analysis. Press Enter for the full definition. and GPT-5.6 choices.
- Release fact
- Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. launch
- Why it matters
- Cursor said Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. improved sustained long-running tasks, complex instruction following and collaboration behavior.
- Release fact
- June roadmap claim
- Why it matters
- At Compile 2026, Cursor said it was training a significantly larger model from scratch with SpaceX.
- Release fact
- June compute claim
- Why it matters
- The Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. post said the new training work used 10x more total compute.
- Release fact
- Forum framing
- Why it matters
- The Compile forum thread described a larger model effort aimed beyond autocomplete and pair programming.
- Release fact
- Current selection
- Why it matters
- Treat those statements as a June roadmap record. Use the current model guide for models that are available now.
| Release fact | Why it matters |
|---|---|
| Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. launch | Cursor said Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. improved sustained long-running tasks, complex instruction following and collaboration behavior. |
| June roadmap claim | At Compile 2026, Cursor said it was training a significantly larger model from scratch with SpaceX. |
| June compute claim | The Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. post said the new training work used 10x more total compute. |
| Forum framing | The Compile forum thread described a larger model effort aimed beyond autocomplete and pair programming. |
| Current selection | Treat those statements as a June roadmap record. Use the current model guide for models that are available now. |
As of August 22, 2026. See the linked Cursor sources for the latest details.
This is covered hands-on in Cursor First Hour — 4 short Units, free to read.
How should a team use Cursor model work?
Knowing what shipped is the easy part. Turning Cursor model work into something a team relies on takes a few deliberate steps, mapped below.
- 1Open the current model guide before choosing a model today.
- 2Use the larger-model announcement as roadmap context for Cursor's long-term differentiation.
- 3Track whether new model availability changes pricing, latency, review quality or agent reliability.
- 4Re-check official model docs before quoting pricing or benchmark claims.
What should stay bounded?
New capability comes with new ways to get it wrong. Keep these boundaries in place as you adopt it.
Do not say the new from-scratch model is generally available unless Cursor publishes that.
Do not mix Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. pricing with a future model's pricing.
Separate benchmark claims from real repo review outcomes.
How does the Cursor harness make models cheaper and better?
A common assumption is that Cursor is a thin desktop client that pipes your prompt straight to a model provider. It is not. Between your prompt and the model sits Cursor's hosted harness - proprietary cloud infrastructure that does context selection, caching, retries, compaction and per-model tuning before and after the call.
Interactive diagram. Tab through its regions; each focused region shows its detail in the panel below.
Each layer the harness adds between your prompt and the model.
Through caching and dynamic context discoveryThe agent pulling only the relevant parts of files, tools and MCP servers into context as needed, instead of loading everything up front. Press Enter for the full definition., the harness reduces agent tokens by around 47% - and that holds even when you are running another company's model.
The same model performs better inside the Cursor harnessCursor's hosted layer around each model (context selection, caching, retries) that makes the same model run better and cheaper than calling it directly. Press Enter for the full definition. than when you call it directly, because Cursor builds a model-specific harness for each model rather than a single generic pipe.
Dynamic context discoveryIntroduced in Cursor 2.4
The older approach loaded everything - every file, every tool definition, all MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. output - into the context window. That bloats context and costs tokens, and as context grows model accuracy degrades. Dynamic context discovery flips this to a filesystem approach: the agent reads from files, tools and MCP output and pulls into the window only what it deems relevant to the task.
- Less irrelevant context means higher accuracy, not just lower cost.
- The token saving compounds with caching to reach the ~47% figure.
- There is no native model memory here - what feels like memory is the harness re-discovering the right context each turn.
Cursor is a harness labCursor as a model-neutral test bed: frontier labs send their newest models to Cursor pre-release to tune them inside the same agent harness every model shares. Press Enter for the full definition.. Model providers send their latest models to Cursor before release so Cursor can test them and build a strong, model-specific harness for each one.
That is also why learning Cursor is a safe bet: new frontier models ship every two to three months, and Cursor stays model-neutral - use whichever frontier model you want behind the same workflow.
Which model should I use for which task?
Start from the current picker, then match the model to the job. High-reasoning choices such as Claude Opus 5 and GPT-5.6 fit planning, auditing and hard bugs. Faster choices such as Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. and Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition. fit implementation and repeated lighter work. The current model guide keeps the live roster, pools and task trade-offs together.
Interactive widget. Tab through its controls; the result updates in the panel below as you change them.
High-reasoning models plan and audit; faster models execute; Auto routes by the selected mode when you do not want to choose.
- Group
- High reasoning
- Use it for
- Plan modeA mode that makes no edits: it researches the codebase and produces an editable plan you review before any code changes. Press Enter for the full definition., planning, auditing, finding bugs, gnarly production issues.
- Examples
- Claude Opus 5, GPT-5.6 Sol, Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition.
- Group
- Fast execution
- Use it for
- Building from a plan, edits across a few files, documentation.
- Examples
- Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition., Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition.
| Group | Use it for | Examples |
|---|---|---|
| High reasoning | Plan modeA mode that makes no edits: it researches the codebase and produces an editable plan you review before any code changes. Press Enter for the full definition., planning, auditing, finding bugs, gnarly production issues. | Claude Opus 5, GPT-5.6 Sol, Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition. |
| Fast execution | Building from a plan, edits across a few files, documentation. | Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition., Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition. |
A brain icon next to a model in the picker marks a thinking/reasoning model; no brain means a standard, faster model.
ComposerCursor's own fast coding model, tuned for the editor and priced well below frontier models; the recommended day-to-day model for executing a plan. Press Enter for the full definition. is Cursor's in-house model, trained specifically on reading and writing code. The rule of thumb at Cursor and outside it: don't use Composer for planning. Use a frontier model to plan, then Composer to execute.
We have a funny quote from my CTO, which is: ComposerCursor's own fast coding model, tuned for the editor and priced well below frontier models; the recommended day-to-day model for executing a plan. Press Enter for the full definition. won't write poetry, it won't do your taxes, but it will write very good code.
Plan with a smart model, build with a fast one
The dominant pattern across field workflows is the same: use a larger, more intelligent model to produce a very detailed plan, then a leaner, faster model to do the building. Once the plan is detailed enough, most frontier models will execute it correctly, so speed becomes the lever - and that is exactly what the fast executors optimize for. Internally Cursor plans this way too, then hands execution to a fast model. You can plan in plan mode, pick the execution model, and click Build to move into the building phase that writes the code.
When a task does heavy MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. tool-calling, don't put a high-reasoning model on it. You'd be burning expensive reasoning tokens while it waits on function feedback.
Reach for an execution model like Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. instead. It keeps spend under control and the loop moves quicker.
A personal per-task cheat sheet
There's no single right answer - model choice is a bit like picking a t-shirt color, so try candidates on your own codebases and build an intuition for where each gets stuck. Keep the evaluation set current rather than copying a picker from an older workshop. A rough set of defaults as of August 14, 2026:
- Planning and high reasoning: Claude Opus 5, GPT-5.6 Sol or Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition..
- Balanced daily work: Claude Sonnet 5Anthropic's medium-tier frontier model in Cursor, with a 200k standard context window and up to 1M context for longer tasks. Press Enter for the full definition. or GPT-5.6 Terra.
- Routine implementation and edits across a few files: Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition..
- Fast, repeated tasks and documentation: Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition..
- Use Claude Fable 5Anthropic's highest-priced generally available Cursor model for long, complex work, with a 30-day retention opt-in for eligible privacy and enterprise accounts. Press Enter for the full definition. for the hardest long-running work only when its higher cost and 30-day Anthropic retention opt-in are acceptable; compare Grok 4.5, Gemini 3.1 ProGoogle's frontier model in Cursor for work that combines images and code, especially UI implementation and visual code analysis. Press Enter for the full definition., the previous Gemini 3.6 Flash and GPT-5.6 Luna when their specific behavior, cost or modality fits.
Judge every model on three axes - performance, cost and speed - and accept that different tasks want different mixes. For planning, most frontier customers are cost-insensitive and just reach for the strongest model; for execution, speed is the lever.
For a gnarly bug or a production-down database error, compare a current Claude Opus 5, GPT-5.6 Sol or Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition. run on the same evidence instead of relying on an older family label.
When you don't want to choose
Two options take the decision off your hands. Auto routes the model per request: since 2026-07-22 it is powered by Cursor Router on Teams and Enterprise plans, which runs a classifier on each agent request and routes it by task type and complexity. What you choose is an optimization mode - Cost, Balance or Intelligence - not a model, and Cost keeps the previous Auto routing logic and its bundled Auto pricing while Balance and Intelligence bill at the routed model's rate. Best-of-N runs the same prompt with the same inputs across multiple models in parallel so you can see for yourself which one fits. You can even run a single model twice to see how much the output varies. There's an upfront cost in running a prompt redundantly, but it's far cheaper than going down a rabbit hole with the wrong model and redoing everything.
One of the most common questions we get at Cursor is: which model should I use for this? Best-of-N is our solution. You can determine it for yourself.
For a high-value debug, some teams push this further into a multi-agent run: kick off several current models on the same problem - for example Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition., Claude Opus 5, Claude Sonnet 5Anthropic's medium-tier frontier model in Cursor, with a 200k standard context window and up to 1M context for longer tasks. Press Enter for the full definition. and GPT-5.6 Sol - all searching the same logs through an MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. like Datadog, then compare their findings. If it costs you a dollar to fully debug the issue, that's worth it.
Which current model pages should I check before choosing?
Use the current model guide first, then open the model-specific docs for availability, pricing and task fit. As of August 14, 2026, the useful current roster is below. It replaces older picker lists on this roadmap page; the June larger-model announcement above stays historical until Cursor publishes a launch.
- Practical read
- Current Grok choice for deeper long-running work.
- Model page
- Grok 4.5
- Practical read
- Keep for controlled comparisons or policies that still depend on it.
- Practical read
- Cursor's coding model for routine implementation and fast execution.
- Practical read
- Balanced Claude choice for daily work.
- Model page
- Claude Opus 5
- Practical read
- High-reasoning Claude choice for difficult planning and debugging.
- Practical read
- Highest-capability long-running work when its higher cost and 30-day Anthropic retention opt-in are acceptable.
- Practical read
- Use when image input, UI work or large-context analysis matters.
- Practical read
- Current speed-tier option for lighter repeated tasks and documentation.
- Model page
- Gemini 3.6 Flash
- Practical read
- Previous Flash version for controlled comparisons and older policies.
- Model page
- GPT-5.6 Luna, Terra and Sol
- Practical read
- Choose the speed, balance or reasoning tier that fits the work.
Current roster checked August 14, 2026. Model docs remain the source of truth for availability and cost; re-check before quoting.
Frequently asked questions
Who is this guide for?
Developers and product teams tracking Cursor model capability and availability.
What should I do next?
Start with one real repo task, capture the prompt and review the result before scaling the workflow.
Is Cursor just a wrapper around other models?
No. Cursor's hosted harness handles context selection, caching, retries and per-model tuning, which cuts agent tokens by about 47% and makes a given model perform better than calling it directly. It also serves Cursor's own Composer models, and providers send their latest models to Cursor pre-release so Cursor can build a model-specific harness for each.
Which model should I use for planning versus building?
Use a current high-reasoning choice such as Claude Opus 5, GPT-5.6 Sol or Grok 4.6 to produce a detailed plan, then a faster executor such as Composer 2.5 to build from it. For tasks with heavy MCP tool-calling, prefer an execution model so you do not burn reasoning tokens waiting on function feedback. If you would rather not choose, Auto routes per request, and Best-of-N runs the same prompt across several models so you can compare. On Teams and Enterprise plans, Cursor Router lets you pick Cost, Balance or Intelligence instead of a model.
Sources & last verified
- Cursor blog: Introducing Composer 2.5
- Cursor model doc: Composer 2.5
- Cursor model doc: Claude Fable 5
- Cursor model doc: Claude Opus 5
- Cursor model doc: Claude Sonnet 5
- Cursor model doc: Gemini 3.1 Pro
- Cursor model doc: Gemini 3.6 Flash
- Cursor model doc: Gemini 3.7 Flash
- Cursor model doc: GPT-5.6 Luna
- Cursor model doc: GPT-5.6 Terra
- Cursor model doc: GPT-5.6 Sol
- Cursor model doc: Grok 4.5
- Cursor model doc: Grok 4.6
- Cursor docs: Models and pricing
- Cursor Community Forum: Compile 2026 announcements
- Cursor product
- Cursor changelog
Cursor ships frequently. Facts verified against primary sources on July 9, 2026.