A programmer works at a laptop displaying AI dashboards and code, with a cat resting nearby.
Grok 4.7 arrived on September 21, 2026, and OpenAI's GPT-6.1 Sol followed on September 29. Most "Grok vs ChatGPT" write-ups on the web are stale, and some of the numbers floating around are wrong. This piece sorts out which model is which, what each costs, where the benchmarks come from, and how a developer can test both without a mess. The comparison matters to anyone choosing a coding assistant, an agent backend or an Azure-hosted model.

A programmer works at a laptop displaying AI dashboards and code, with a cat resting nearby. Which models are actually being compared​

"ChatGPT" is not one model this fall. It's a family with several names, and getting them straight is half the battle.

  • Grok 4.7: xAI's model (now branded SpaceXAI on some pages) is the fastest of the pair on output speed in Artificial Analysis' release comparison. xAI's launch post says it uses a new, larger base model than Grok 4.6 — wait, that detail is from xAI's own announcement rather than the search results; xAI describes a longer reinforcement-learning run weighted toward tasks that take many hours, and says the model is better at verifying its own work and managing longer context.
  • GPT-6 Astra: OpenAI's top model. Its announcement says it began rolling out on September 3, 2026. OpenAI lists availability to ChatGPT Plus, Pro, Business and Enterprise users, plus the OpenAI API, Microsoft Azure and AWS Bedrock.
  • GPT-6.1 Sol: OpenAI describes it as an upgrade to GPT-6 Sol that nearly matches Astra on agentic coding, computer use and professional work, at roughly one-fifth of Astra's token prices.
  • GPT-6 Luna: a cheaper everyday model, listed by OpenAI at $0.10 per million input tokens and $0.50 per million output tokens.

One correction to a popular framing: GPT-6.1 Sol is not just an "API-only" model. OpenAI says it is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.1-sol. It is not yet available in Chat. Claims about which model powers a given consumer plan are therefore shaky, and this article doesn't rely on them.

Specs and pricing side by side​

All figures below come from vendor documentation unless noted.

SpecGrok 4.7GPT-6.1 SolGPT-6 Astra
ReleaseSep 21, 2026Sep 29, 2026Sep 3, 2026 (rollout)
Context window500,000 tokens1.05M tokens1.05M tokens
Input price per 1M tokens$2$2$10
Output price per 1M tokens$6$10$50
Cached input per 1M tokens$0.50$0.10$1.00
InputsText, imageText, image (per Artificial Analysis)Text, image (per Artificial Analysis)
OutputTextTextText

Some details worth knowing:

  • Context: xAI's developer guide lists 500,000 tokens for Grok 4.7. OpenAI's API model pages list about 1.05 million tokens and a 128,000-token maximum output for both Astra and Sol. Claims that OpenAI hasn't disclosed context windows are outdated.
  • Long prompts: OpenAI says GPT-6.1 Sol prompts above 272,000 input tokens are billed at twice the input and cache rates and 1.5 times the output rate for the full request. I could not confirm from xAI's official materials the specific long-context rate card circulating in some write-ups, so check xAI's pricing page before budgeting around 200K-plus prompts.
  • Fast modes: xAI's docs say Grok 4.7 Fast costs twice the standard rate and is currently available only in Cursor and Grok Build, not on the public xAI API. OpenAI offers a Fast mode for Astra at double the price for up to twice the speed.
  • Reasoning controls: Grok 4.7 offers low, medium, high (the default) and xhigh. Tools listed include function calling, web search, X search and code execution.

What the benchmarks say, and who ran them​

Benchmarks come from two kinds of sources here: vendors grading themselves, and Artificial Analysis running its own tests. Don't mix them.

xAI's self-published figures for Grok 4.7:

  • Terminal-Bench 4.0: 37.6%, versus 20.3% for Grok 4.6. In the same table xAI lists 57.9% for Fable 5.1 Max, which is another vendor's model, not GPT-6 Astra.
  • CursorBench 4.0: 46.3%.
  • DeepSWE v1.1: 71.0% (high effort).
  • EEBench: 64.0%.
  • AA Briefcase v1.1: 1,657.
  • Harvey Legal Agent Benchmark: 19.6%, which xAI's table still shows ahead of the other models it lists.
  • HackerBench v0.3: xAI says it lets through only 3.3% of risky dual-use cyber prompts.

That last point corrects a common misreading. 96.7% is the complement of 3.3%, and it is xAI's own safety measure, not independent evidence.

Artificial Analysis' independent runs: On its Intelligence Index (v4.3.2, ten evaluations), the Grok 4.7 versus GPT-6 Astra (Xhigh) comparison shows GPT-6.1 Sol (Max) at 52 against Grok 4.7 (Xhigh) at 46 in the release comparison. A separate page puts GPT-6.1 Sol (High) at 50 against Grok 4.7 (High) at 46. Its Astra comparison page lists 52 for Astra (Xhigh) and 46 for Grok.

On Terminal-Bench 4.0, Artificial Analysis' own runs differ from the vendor numbers: Sol (High) at 52% against Grok 4.7 (High) at 25%, and its Astra page shows 60% versus 26% for the Xhigh configurations. The gap is real in both datasets. The absolute numbers aren't comparable, though, because the harnesses differ. The popular "57.9% for Astra" figure is not supported by the vendor material I reviewed.

Grok isn't beaten everywhere. In the Artificial Analysis High-versus-High comparison, Grok 4.7 scores 1631 on AA-Briefcase against 1465 for Sol, and 1705 against 1509 on GDPval-AA. Knowledge-work tasks are where xAI's pitch is strongest.

The cost per task twist​

Per-token prices tell only part of the story. Artificial Analysis publishes cost per task, and the results flip the naive reading:

  • Sol (High) versus Grok 4.7 (High): $0.32 against $2.73 per task, with Sol using about 13k output tokens per task against 66k for Grok.
  • Astra (Xhigh) versus Grok 4.7 (Xhigh): $2.31 against $3.74 per task, with Astra using about 17k output tokens against 81k.

In other words, the model with the pricier output rate can come out cheaper because it spends far fewer reasoning tokens. Grok does have the speed advantage: 84 tokens per second for Grok 4.7 (Xhigh) against 54 for GPT-6.1 Sol (Max). Artificial Analysis' Astra-versus-Grok page, meanwhile, shows Astra's time per task at about 315 seconds against Grok's roughly 937.

These are one tester's runs, with specific configurations and a specific index mix. Your prompts will behave differently. That's the argument for measuring on your own workload.

Microsoft and enterprise angle​

This isn't an Azure-integration story for Grok or Sol specifically, but there are two concrete hooks:

  • OpenAI lists Microsoft Azure as an availability channel for GPT-6 Astra, alongside the OpenAI API and AWS Bedrock. Enterprise administrators can enable Astra for their workspace, and it is off by default at launch.
  • OpenAI also notes that Astra's added safety checks can pause or stop legitimate work, including defensive cybersecurity. In ChatGPT or Codex you may be asked to review an action. In the API, the task simply stops. Plan for that in automation.

Nothing in the evidence I reviewed shows a special integration of Grok 4.7 or GPT-6.1 Sol with Entra, Intune or Microsoft 365, so don't assume one.

Testing Grok 4.7 alongside an OpenAI-based app​

xAI's API is designed to be reachable from OpenAI-style client libraries, which makes a side-by-side test cheap. Treat it as a starting point, not a guaranteed drop-in.

  1. Create an xAI API key. It is separate from your OpenAI key.
  2. Point your client at [url]https://api.x.ai/v1[/url]. xAI also offers a US regional endpoint at [url]https://us.api.x.ai/v1[/url], priced at a 10% premium.
  3. Set the model name to grok-4.7.
  4. Note that xAI's current developer guide centres on the Responses API. Check which endpoint and features your code actually uses.
  5. Set a prompt_cache_key (or the x-grok-conv-id header on Chat Completions). xAI says this makes cache hits reliable. Without it you may pay full input price on a cache-cold server.
  6. Run the same prompt suite against both models and record completion rate, retries, latency, tool-call behavior and real token spend, not just list prices.
  7. Re-check reasoning-effort defaults. Grok defaults to high, and effort levels change both cost and quality.

A minimal sketch, based on xAI's own JavaScript example, uses the OpenAI client with baseURL set to the xAI endpoint and calls client.responses.create with model: 'grok-4.7'.

What to take away​

  • Pick Sol or Astra for agentic coding: both vendor and independent data show a clear Terminal-Bench lead for OpenAI's models.
  • Pick Grok 4.7 for speed and cheap output: it generates faster, costs less per output token, and scores well on knowledge-work benchmarks like GDPval-AA.
  • Pick Sol for huge repos and long documents: its 1.05M window is double Grok's 500K.
  • Measure cost per task, not per token: the token-heavy reasoning style of Grok 4.7 narrows or erases its price advantage in Artificial Analysis' runs.
  • Treat leaderboards as snapshots: Artificial Analysis shows different values depending on configuration (High, Xhigh, Max), so cite the configuration along with the number.

A fair verdict: OpenAI leads on measured capability and, per Artificial Analysis, often on cost per completed task. Grok 4.7 earns its keep on throughput and output pricing. The honest answer to "which is better" is to run both on your actual workload for a couple of weeks before committing.

 

References

  1. Grok 4.7 and the GPT-6 Family: What Changed This Fall - shattered.io shattered.io 2026-10-19T12:32:52+00:00
  2. GPT-6.1 Sol openai.com
  3. GPT-6 Astra: A new generation of intelligence openai.com