Grok 4.6 Targets Long-Running Agents With Fewer Tool Turns and Flat API Pricing
SpaceXAI's Grok 4.6 improves long-horizon agent benchmarks, retains $2/$6 pricing, and offers a 500,000-token API context.
SpaceXAI released Grok 4.6 on August 12, positioning the model for long-running agents that research topics, operate across codebases, use tools, and refine work over multiple steps. The release improves substantially over Grok 4.5 on several agent evaluations while retaining headline API prices of $2 per million input tokens and $6 per million output tokens.
The most distinctive result comes from Artificial Analysis’s AA-Briefcase benchmark for long-horizon knowledge work. Grok 4.6 completed its tasks in approximately 53 turns and consumed about 500 million input tokens across the evaluation, compared with roughly 103 turns and two billion input tokens for Claude Opus 5 at maximum effort. The comparison supports a specific efficiency claim: on this benchmark, Grok required about half as many turns and one-quarter as many input tokens.
That finding is narrower than the original social-media description that Grok 4.6 uses half the turns of “other leading models.” Artificial Analysis published the detailed comparison against Claude Opus 5 Max, not every competing frontier model. It is nevertheless consequential because long agent loops repeatedly resend or accumulate context, making the number of steps a direct component of latency and cost.
1. Grok 4.6 Extends the Grok 4.5 Training Line
SpaceXAI describes Grok 4.6 as a continuation of Grok 4.5 rather than a newly disclosed architecture. The company has not published a parameter count, training-compute figure, or detailed architectural specification for the release.
The model received a longer supplemental training run than Grok 4.5. SpaceXAI says this stage used curated model-generated material for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer and training recipe.
Grok 4.5 was then used to regenerate supervised fine-tuning trajectories across multiple reasoning-effort settings, agent harnesses, and domains including STEM, software engineering, and knowledge work. Model-based checks filtered problematic trajectories before the resulting checkpoint proceeded through reinforcement learning.
The reinforcement-learning environments covered general coding and knowledge work as well as more specialized tasks involving kernel optimization, web development, and computer-aided design. SpaceXAI says longer test trajectories also produced more instances of the model testing and verifying its own work before continuing.
Cursor data belongs to this model lineage, but the available disclosures require careful wording. Cursor said Grok 4.5 was jointly trained with SpaceXAI as a mixture-of-experts model using trillions of tokens of Cursor data. That material captured both codebase interactions and exchanges between developers, agents, and software tools.
The Grok 4.6 announcement says the new model builds on Grok 4.5 and received additional engineering data, but it does not quantify how much new Cursor data was added specifically for version 4.6. Describing Grok 4.6 as benefiting from a Cursor-trained foundation is supported; claiming a newly disclosed volume of Cursor data for the 4.6 supplemental run is not.
2. Agent Benchmarks Show Large Gains, but Not a Clean Sweep
Grok 4.6 scored 61 on version 4.1.1 of the Artificial Analysis Intelligence Index, five points above Grok 4.5 and equal to GPT-5.6 Sol Max in SpaceXAI’s launch comparison. Fable 5 Max scored 62. The index combines nine evaluations covering agentic work, terminal tasks, science, general knowledge, and reasoning.
The release’s clearest improvements appear on agent-oriented tests:
| Evaluation | Grok 4.6 High | Grok 4.5 High | GPT-5.6 Sol Max | Fable 5 Max |
|---|---|---|---|---|
| GDPVal-AA v2 | 1,753 | 1,526 | 1,728 | 1,741 |
| CursorBench v3.2 | 69.9% | 66.7% | 67.2% | 70.5% |
| DeepSWE v1.1 | 65.9% | 54.0% | 73.0% | 70.0% |
| FrontierCode v1.1 Extended | 61.3% | 56.6% | 60.6% | 63.6% |
| APEX-Agents | 57.5% | 47.1% | 56.7% | 59.2% |
| Terminal-Bench v3.0 | 26.0% | 15.7% | 34.6% | 34.1% |
| AA-Briefcase | 1,577 | 1,313 | 1,502 | 1,574 |
GDPVal-AA v2 evaluates economically valuable professional tasks. Artificial Analysis independently reported that Grok 4.6’s result was behind only Claude Opus 5 and statistically indistinguishable from Fable 5 and Qwen3.8 Max because their confidence intervals overlapped.
AA-Briefcase produced a similar quality result with a more unusual efficiency profile. Grok 4.6 reached an Elo rating of 1,577, slightly above the 1,574 shown for Fable 5 Max and well above GPT-5.6 Sol Max’s 1,502 in SpaceXAI’s table. Artificial Analysis characterized Grok’s output as consistently competitive across rubric compliance, presentation, and analytical quality.
The results do not establish Grok 4.6 as the strongest model on every agent workload. GPT-5.6 Sol Max led it by 7.1 percentage points on DeepSWE v1.1 and by 8.6 points on Terminal-Bench v3.0. Fable 5 Max also remained ahead on CursorBench, FrontierCode, APEX-Agents, Terminal-Bench, and APEX-SWE in the launch table.
Benchmark conditions differ by harness and reasoning setting, and SpaceXAI notes that competing figures were selected from vendors’ published system cards or public leaderboards. The table is therefore evidence of competitive performance across several environments, not a controlled declaration that one model dominates all agent systems.
3. Lower Turn Counts Change the Cost Calculation
Artificial Analysis measured an average cost of $0.84 per task while evaluating Grok 4.6 on its Intelligence Index. It placed the model on the cost-performance frontier for each agentic evaluation included in that index.
The economic advantage comes from two separate factors. Grok 4.6 has lower published token prices than several nearby frontier models, and its AA-Briefcase runs accumulated less context because they used fewer turns. A long-running agent may send instructions, conversation history, retrieved documents, and tool results back to the model after every action. Reducing the number of cycles can therefore lower input-token consumption even when the final answer is similar.
This does not mean every Grok 4.6 application will cost $0.84 or use half as many turns. Those figures describe Artificial Analysis’s evaluation configuration. Real costs depend on reasoning effort, tool design, prompt caching, context growth, retries, and whether the agent completes the task successfully.
Raw speed is another limitation. Artificial Analysis measured approximately 62 output tokens per second and a 44.82-second time to first token through SpaceXAI’s API, both slower than the comparison medians shown on its model page. Turn efficiency can reduce the duration of an entire workflow without making each individual response fast.
4. API and Cursor Deployments Have Different Limits
The native SpaceXAI API uses the model identifier grok-4.6. It accepts text and images, returns text, and supports low, medium, high, and xhigh reasoning effort, with high as the default. Its documented tools include function calling, web and X search, and code execution.
SpaceXAI documents a 500,000-token context window, a February 1, 2026 knowledge cutoff, and no stated text-output limit. Below 200,000 prompt tokens, API pricing is $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. Those rates double to $4, $1, and $12 once a prompt exceeds 200,000 tokens.
The company recommends assigning a stable prompt-cache key to a conversation so successive requests reach the same server and can reuse cached context. It also recommends context compaction for long agent loops. Both mechanisms matter because the advertised context capacity does not remove the cost of repeatedly processing a growing history.
Cursor’s deployment exposes a smaller documented context window of 256,000 tokens. It gives Grok 4.6 access to Cursor’s file search, web access, file reading and editing, terminal, browser, image-generation, and rules-retrieval tools. Standard on-demand rates match the lower API tier, while Cursor’s Fast option costs $4 per million input tokens, $1 per million cached input tokens, and $12 per million output tokens.
At launch, Grok 4.6 was available through Cursor, Grok Build, the SpaceXAI API, OpenRouter, Vercel, and Cloudflare. Cursor and Grok Build offered double included usage for the first week, but that was a time-limited launch promotion rather than a permanent pricing term.
Frequently Asked Questions
When was Grok 4.6 released?
SpaceXAI released Grok 4.6 on August 12, 2026.
Does Grok 4.6 always use half as many agent turns?
No. The approximately 53-versus-103-turn comparison comes specifically from Artificial Analysis’s AA-Briefcase evaluation against Claude Opus 5 Max.
Was Grok 4.6 trained on Cursor data?
Its Grok 4.5 foundation was trained with trillions of tokens of Cursor data. SpaceXAI has not disclosed how much additional Cursor data, if any, was used specifically in the Grok 4.6 supplemental run.
What does the Grok 4.6 API cost?
Below 200,000 prompt tokens, the documented rates are $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. The rates double above that threshold.
Does Grok 4.6 have the same context window in Cursor and the API?
No. SpaceXAI documents 500,000 tokens for its API, while Cursor documents a 256,000-token context window for its deployment.
Sources
Share