Grok 4.6 catches GPT-5.6 Sol. The price may matter even more
Grok 4.6 ties GPT-5.6 Sol on a composite intelligence index, starts at $2 per million input tokens and focuses on long-running agents. We examine what xAI's launch actually proves.
Grok 4.6 is more than a routine version bump. xAI's new model has entered the top group of public evaluations, tied GPT-5.6 Sol on the Artificial Analysis composite index and retained the same base price as Grok 4.5. A few weeks ago, this combination of benchmark scores and a $2-per-million-token input rate would have sounded unlikely.
The more important change is how the model is meant to work. xAI describes Grok 4.6 as a system for long-running agent tasks: researching a topic, analyzing multiple sources, navigating a repository and building a complete application across dozens of steps. It is not merely a chatbot expected to produce a better single response. It is supposed to retain the goal, use tools, verify its own output and revisit a problem when its first solution fails.
The model launched on August 12, 2026. It is available through the API, Grok Build, Cursor and partners including OpenRouter, Vercel and Cloudflare.
Grok 4.6 at a glance
| Feature | Grok 4.6 |
|---|---|
| API model ID | grok-4.6 |
| Launch date | August 12, 2026 |
| Modalities | text and image input, text output |
| Context window | 500,000 tokens |
| Price below 200K prompt tokens | $2 input, $0.50 cached input, $6 output / 1M tokens |
| Price at or above 200K prompt tokens | $4 input, $1 cached input, $12 output / 1M tokens |
| Reasoning levels | low, medium, high and xhigh |
| Default reasoning level | high |
| Knowledge cutoff | February 1, 2026 |
| Primary use cases | coding, agents, knowledge work, interactive and visual projects |
The long-context threshold deserves attention. The advertised $2/$6 pricing applies while the prompt remains below 200,000 tokens. Once it reaches that threshold, the higher rate applies to every token in the request, not only the portion above the line. A very large repository or document set can therefore produce a step change in cost.
xAI also offers a faster variant at twice the base price. “Fast” does not imply a separate quality tier. It is primarily an option for workloads where latency matters more than inference cost.
What changed from Grok 4.5
Grok 4.5 arrived only a month earlier and had already pushed xAI toward coding and agentic work. Version 4.6 develops that direction instead of trying to become a completely different product.
According to xAI's training description, the new model received a longer supplemental training run. The company used model-generated reasoning and advanced technical data, high-quality engineering data, an improved optimizer and a revised training recipe. Grok 4.5 then regenerated SFT trajectories across reasoning levels, agent harnesses and domains including STEM, software engineering and knowledge work.
Reinforcement learning covered general coding and domain-specific environments for kernel optimization, web development and computer-aided design. That helps explain why the announcement emphasizes completing a project rather than producing a single impressive answer.
The most practical claim is that Grok 4.6 performs more self-testing and verification during long trajectories. This matters because coding agents rarely fail for lack of syntax knowledge. They fail after losing the original objective, overlooking side effects or declaring the work complete before testing it.
Benchmarks: tied with GPT-5.6 Sol, but not on every task
The headline result is 61 on the Artificial Analysis Intelligence Index. GPT-5.6 Sol has the same score, while Fable 5 Max reaches 62. Axios framed the result as SpaceXAI returning to the immediate frontier group.
A composite score cannot tell a team which model fits its workflow. The detailed results are more informative:
| Evaluation | Grok 4.6 High | Grok 4.5 High | GPT-5.6 Sol Max | Fable 5 Max |
|---|---|---|---|---|
| AA Intelligence Index | 61 | 56 | 61 | 62 |
| GDPVal-AA v2 | 1753 | 1526 | 1728 | 1741 |
| CursorBench 3.2 | 69.9% | 66.7% | 67.2% | 70.5% |
| DeepSWE 1.1 | 65.9% | 54.0% | 73.0% | 70.0% |
| FrontierCode 1.1 Extended | 61.3% | 56.6% | 60.6% | 63.6% |
| APEX-Agents | 57.5% | 47.1% | 56.7% | 59.2% |
| Terminal-Bench 3.0 | 26.0% | 15.7% | 34.6% | 34.1% |
| AA-Briefcase | 1577 | 1313 | 1502 | 1574 |
There is no universal winner in that table. Grok 4.6 beats GPT-5.6 Sol on CursorBench, FrontierCode, APEX-Agents and the long-horizon AA-Briefcase knowledge-work evaluation. OpenAI's model has a large lead on DeepSWE and Terminal-Bench, both relevant to difficult software and terminal work. Fable 5 remains consistently strong and leads several evaluations, although it is considerably more expensive.
The right conclusion is not “Grok is better than GPT.” A more defensible reading is that Grok 4.6 belongs in the same performance class, with a different strength profile and a substantially lower base API price.
Why agent benchmarks require caution
A traditional evaluation gives a model a question and scores its answer. An agent evaluation depends on many more components: the system prompt, tools, context management, retry policy, history compaction and verification loop.
That distinction is particularly important for CursorBench. Grok has been developed in close collaboration with Cursor, and its training included trajectories from multiple agent harnesses. The score remains useful, but it would be misleading to present it as a pure measure of raw model intelligence. It evaluates an operating system made from a model and its environment.
xAI also notes that competitor scores are the best self-reported or publicly available results. Models may be running at different effort settings and with different tool stacks. Before migrating a production workflow, teams should build a private suite of 20 to 50 repeatable tasks and measure:
- the share completed without manual correction;
- tool calls, backtracking and retries;
- total task cost rather than the price per million tokens;
- time to an output that passes tests and review;
- stability across repeated runs.
Price may be Grok 4.6's largest advantage
Grok 4.6 keeps Grok 4.5's base rate: $2 per million input tokens and $6 per million output tokens. Cached input costs $0.50 per million, which can matter for agents that repeatedly inspect the same files and context.
A cheap token does not guarantee a cheap task. A model may take more steps, generate longer reasoning or repeat an operation several times. In practice, agent economics look more like:
model price × context growth × turns × recovery cost
Grok 4.6 is interesting because the public results point not only to inexpensive tokens, but also to competitive long-horizon task performance. If it needs a similar or smaller number of turns than more expensive rivals, the saving becomes real. If engineers repeatedly repair its output, the advantage disappears.
The long-context tier reinforces this point. Requests that reach 200,000 prompt tokens are billed at $4 input and $12 output per million. Teams should not fill the 500K window simply because it exists. Retrieval, a concise repository map and periodic state summaries can be cheaper and easier for the model to navigate.
Where Grok 4.6 makes the most sense
Based on the specification and early evaluations, the model is worth testing in four areas.
- Large repository changes. A 500K context window and long-trajectory training fit migrations, cross-file work and tasks with several test cycles.
- Building a first application version. xAI presents Grok as able to move from a broad idea to a working interface and refine it through feedback.
- Knowledge work. GDPVal-AA and AA-Briefcase suggest potential for reports, analysis and complex professional deliverables.
- Cost-sensitive agents. The price can support more verification, more frequent agent runs or a larger workload within the same budget.
It is not automatically the best choice for every terminal-heavy assignment. Terminal-Bench indicates that GPT-5.6 Sol and Fable 5 retain a clear advantage in some hard scenarios. Nor does the model know events after February 1, 2026 by default. Current information requires Web Search or X Search, whose tool calls are billed separately.
Availability and integration
Developers can call the model as grok-4.6. It supports function calling, structured outputs and configurable reasoning. It accepts text and images and returns text.
Beyond the API, the model is available in Grok Build and Cursor. xAI also names OpenRouter, Vercel and Cloudflare as launch partners. Grok Build and Cursor include twice the normal usage during the first week, a temporary launch promotion rather than a permanent pricing feature.
For teams moving from Grok 4.5, the unchanged base rates are welcome. The migration still deserves an evaluation at each reasoning_effort level. high is the default, but routine classification and simple edits may be cheaper at low or medium, while the hardest tasks may benefit from xhigh.
The API documentation lists 150 requests per second and 50 million tokens per minute for the model, although actual account limits and regional capacity can vary. It is currently listed in us-east-1 and us-west-2 clusters.
Does Grok 4.6 change the model race?
Yes, but not because a single leaderboard settles the competition. Grok 4.6 shows that a model priced at $2/$6 can enter the same conversation as GPT-5.6 Sol and Fable 5. More importantly, xAI is moving beyond Grok's reputation for entertaining responses and optimizing for work that can be inspected: code, a document, a report or a functioning application.
The fairest assessment today is that Grok 4.6 is a serious option for long-running agents and cost-conscious teams, but it does not automatically replace GPT-5.6 Sol or Fable 5 across every workload. Its weaker terminal results matter. Its Cursor coding, knowledge-work performance and pricing also matter.
That cost pressure may be the launch's most important effect. When several models operate near the same capability level, the winner is not the one with the tallest bar on a single slide. It is the one that completes real work with an acceptable bill and an acceptable amount of human cleanup.
(Photo: Farzad Nazifi, CC0 1.0 / Wikimedia Commons. Crop by TreffikAI.)
