xAI shipped Grok 4.7 on September 21, 2026, and led with the same number that made Grok 4.6 a talking point among developers: $2 per million input tokens and $6 per million output tokens. That pricing, confirmed by xAI’s own release notes and repeated across coverage from Kingy AI, LLM Stats, and ai-tldr.dev, undercuts both OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5.1 by a wide margin. But the benchmark scores that followed within 24 hours told a messier story: on the Artificial Analysis Intelligence Index, Grok 4.7 landed at 46, seven points behind the 53 that both GPT-6 Astra and Claude Fable 5.1 posted, according to reporting from AI Weekly and officechai.com.
That gap is the story. xAI built Grok 4.7 to be the cheapest frontier-class model on the market, and by that measure it succeeded. Whether it is a frontier model at all, once price is stripped out of the comparison, is the question now dividing the outlets that tested it against GPT-6 Astra and Claude Fable 5.1.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Grok 4.7 Costs: The $2/$6 Pricing Breakdown
xAI kept Grok 4.7’s standard API pricing identical to Grok 4.6: $2 per million input tokens and $6 per million output tokens for prompts under 200,000 tokens, with cached input billed at $0.50 per million tokens, a structure laid out in xAI’s own SpaceXAI developer documentation published in September 2026. Cross the 200,000-token threshold and the rate roughly doubles, moving to $4 per million input tokens, $1 per million cached tokens, and $12 per million output tokens, a structure confirmed by both Kingy AI and Mitrade’s live-news desk. xAI also quietly introduced a faster variant, Grok 4.7 Fast, which SpaceXAI’s release notes and ComputingForGeeks both price at exactly double the standard tier, $4 per million input tokens and $12 per million output tokens, for teams trading a few points of latency for throughput.
The context window carries over from Grok 4.6 at 500,000 tokens, a figure confirmed both by Mitrade and by xAI’s own SpaceXAI developer documentation published in September 2026. What changed under the hood is less clear from public disclosures, though AI Weekly’s release coverage cites a roughly 40% increase in parameter count with pricing held flat, a move xAI has now repeated across two consecutive releases rather than pass rising compute costs on to developers.
Grok 4.7 Pricing vs GPT-5.6 Sol vs Claude Fable 5.1
Lined up against the two other flagship API products on the market, the price gap is the widest signal in this launch. According to figures compiled by officechai.com, GPT-5.6 Sol charges $4 per million input tokens and $20 per million output tokens, while Claude Fable 5.1 charges $10 per million input tokens and $50 per million output tokens. Grok 4.7’s $2/$6 rate makes it roughly half the input cost of GPT-5.6 Sol and a fifth of Claude Fable 5.1’s, while its output rate runs at about a third of Sol’s and close to an eighth of Fable 5.1’s. Even xAI’s speed-optimized Grok 4.7 Fast tier, priced at $4 input and $12 output per million tokens according to ComputingForGeeks’ September 2026 coverage, still undercuts GPT-5.6 Sol’s output rate and comes in well below Claude Fable 5.1 on both ends.
| Model | Input ($/M tokens) | Output ($/M tokens) | Cached input ($/M tokens) | Context window |
|---|---|---|---|---|
| Grok 4.7 (standard) | $2.00 | $6.00 | $0.50 | 500,000 tokens |
| Grok 4.7 (200K+ prompts) | $4.00 | $12.00 | $1.00 | 500,000 tokens |
| GPT-5.6 Sol | $4.00 | $20.00 | n/a in cited data | n/a in cited data |
| Claude Fable 5.1 | $10.00 | $50.00 | n/a in cited data | n/a in cited data |
Anyone running high-volume agentic workflows, where output tokens pile up fast across multi-step tool calls, will feel that output-price gap more than the input-price gap. It is the same dynamic that made Claude Opus 5.5’s cost-to-performance ratio a talking point earlier this year: cheaper output pricing changes which model teams default to for long-running agents, even when the flagship model scores higher on paper.
The Benchmark Reality: Where Grok 4.7 Falls Behind
The Artificial Analysis Intelligence Index is the aggregate score that gets cited most often when outlets want a single number to compare frontier models, and it is where Grok 4.7’s gap shows up most plainly. officechai.com’s headline on the release put it directly: Grok 4.7’s score jumped two points over Grok 4.6, moving it up the board, but it still landed below GPT-5.6 Sol, Muse Spark 1.3, and Claude Fable 5.1. AI Weekly and orcarouter.ai both independently reported the same 46-point score for Grok 4.7 against 53 for both GPT-6 Astra and Claude Fable 5.1 on that index, a seven-point deficit that works out to roughly 13% lower than the leading pack.
That single-number gap is the clearest read on why “bargain frontier model” and “frontier model” are not the same claim. xAI priced Grok 4.7 like a budget option and it is testing like one on the aggregate index, even while it wins specific categories outright.
AA-Briefcase: Closer Than the Headline Number Suggests
Not every aggregate measure shows the same spread. On AA-Briefcase, a benchmark that scores models on realistic professional work tasks using an Elo-style rating, Grok 4.7 posted 1657, up 111 points from Grok 4.6, according to 247wallst.com. That result places Grok 4.7 just behind Claude Opus 5 and Claude Fable 5.1 rather than far behind them, a much tighter margin than the Intelligence Index implies. The gain over its own predecessor is the more interesting number here: a 111-point jump in one release cycle is a larger single-generation leap than xAI’s pricing announcement alone would suggest.
Where Grok 4.7 Actually Wins: Coding and Legal Benchmarks
Benchmark results from officechai.com’s detailed testing round show a more favorable picture for xAI once the comparison narrows to specific coding and domain tasks rather than a blended intelligence score. On DeepSWE v1.1, a software-engineering benchmark, Grok 4.7 scored 71.0%, ahead of GPT-5.6 Sol’s 65.2% and roughly in line with Claude Fable 5.1’s 70.0%, though it trailed the Anthropic comparison model tested at 72.7%. On the Harvey Legal Agent Benchmark, Grok 4.7 posted 19.6%, ahead of GPT-5.6 Sol’s 15.8%, Claude Fable 5.1’s 6.7%, and a separate Anthropic model’s 2.5% — a category where Mitrade’s coverage led with the framing that Grok 4.7 “outscores Anthropic and OpenAI models on legal agent work for less.”
Terminal-Bench 4.0, which tests command-line and agentic tool use, produced a split result: Grok 4.7 scored 38.0%, well ahead of GPT-5.6 Sol’s 20.3% and roughly matching an Anthropic comparison model at 37.3%, but Claude Fable 5.1 led the field at 57.9%. That inconsistency across benchmarks, strong against one competitor and weak against another on the same test, is the pattern that keeps showing up across this release. There is no single model that Grok 4.7 beats across the board, and no single model that beats Grok 4.7 across the board either.
| Benchmark | Grok 4.7 | GPT-5.6 Sol | Claude Fable 5.1 | Anthropic model (separate test) |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 46 | n/a in cited data | 53 | n/a in cited data |
| DeepSWE v1.1 | 71.0% | 65.2% | 70.0% | 72.7% |
| Terminal-Bench 4.0 | 38.0% | 20.3% | 57.9% | 37.3% |
| Harvey Legal Agent Benchmark | 19.6% | 15.8% | 6.7% | 2.5% |
| CursorBench 4.0 (xhigh effort) | 46.3% | 40.4% | 41.7% | 46.6% (Claude Opus 5 max) |
CursorBench and the Cost-Per-Task Number That Matters More
Kingy AI’s benchmark tracking added a detail that the raw accuracy scores miss: cost per completed task. On CursorBench 4.0, Claude Opus 5 running at its maximum effort setting scored 46.6% at an average cost of $11.95 per task. Grok 4.7 running at its highest effort setting scored 46.3%, a near-identical result, for $6.01 per task, roughly half the price. Dropping to Grok 4.7’s “high” effort setting brought the score down slightly to 43.9% but cut the cost further to $4.69 per task.
That comparison is probably the most useful number in this entire release for anyone actually deciding which API to call. A 0.3-point accuracy gap for half the cost is a straightforward trade for most production workloads. It is a very different framing from the Intelligence Index’s seven-point gap, and it explains why coverage of this launch has split between “Grok 4.7 lags the frontier” and “Grok 4.7 undercuts the frontier for nearly the same result,” depending on which benchmark the outlet leads with.
Grok 4.7 vs Grok 4.6: What Actually Changed
xAI’s own framing treats Grok 4.7 as an iteration on Grok 4.6 rather than a ground-up rebuild, and the numbers back that up. Pricing did not move. The 500,000-token context window did not move, per Mitrade’s reporting. What moved was the Intelligence Index score, up two points, the AA-Briefcase Elo rating, up 111 points, and the DeepSWE score, up from 65.2%-range territory for the prior generation to 71.0%. AI Weekly’s release alert also cited a roughly 40% parameter increase behind the model, without xAI publishing exact parameter counts.
That is a meaningfully faster cadence than either OpenAI or Anthropic has kept in 2026. xAI now has two model releases inside about a month, per multiple outlets’ framing of Grok 4.7 landing roughly one month after Grok 4.6 shipped. Whether that pace produces compounding gains or just incremental point-chasing on specific benchmarks is the open question heading into whatever comes after Grok 4.7. It is worth noting Elon Musk has separately discussed scaling Grok’s parameter count well beyond this release, a track covered in earlier reporting on xAI’s roadmap toward larger models, though that reporting is distinct from the Grok 4.7 launch itself.
Historical Context: xAI’s Pricing Strategy Since Grok 3
xAI has run the same playbook since it entered the paid API market: ship a model priced meaningfully below OpenAI and Anthropic’s published rates, then let benchmark scores catch up over subsequent point releases rather than leading with them. Grok 4.6 introduced the $2/$6 standard tier and the 200,000-token long-context threshold that Grok 4.7 inherited without alteration. That continuity matters commercially: developers who built pricing assumptions into production systems around Grok 4.6 do not need to re-architect anything to adopt Grok 4.7, since the API contract, rate limits, and cost model carried over unchanged.
Contrast that with how OpenAI and Anthropic have typically priced flagship releases, generally holding or raising rates on new flagship tiers while offering cheaper “mini” or distilled variants alongside them. xAI’s approach collapses that tiering: there is no separate budget Grok model sitting below Grok 4.7. The flagship model is the budget option, which is a different market position than either competitor currently occupies, and it’s part of why comparisons to Muse Spark’s own price-performance positioning against GPT-6 Astra and Claude Fable keep surfacing in the same conversation.
Market Impact: What Cheap Frontier Pricing Means for Enterprises
For engineering teams running high-token-volume workloads, agentic pipelines, batch document processing, long-context retrieval, the arithmetic favors Grok 4.7 even accounting for its lower aggregate benchmark score. A team processing tens of millions of output tokens a month sees a direct, linear cost reduction by switching from Claude Fable 5.1’s $50-per-million output rate to Grok 4.7’s $6, a savings large enough to absorb a few percentage points of accuracy loss on tasks where Grok 4.7 already trails.
The picture inverts for teams building on tasks where the Intelligence Index gap actually shows up in output quality, legal review, complex multi-step reasoning, or anything where a wrong answer carries real cost. There, the seven-point Intelligence Index deficit and the mixed Terminal-Bench result are reasons to stick with a pricier flagship rather than chase the discount. No single outlet in this release cycle has published data suggesting Grok 4.7 is a strict upgrade over Claude Fable 5.1 or GPT-6 Astra on quality alone; every comparison published so far pairs a price advantage with a capability trade-off somewhere in the benchmark suite.
Competitive Landscape: The Three-Way API Pricing Race
Grok 4.7’s launch lands in the middle of an active pricing and benchmark contest between xAI, OpenAI, and Anthropic that has played out in public release notes for most of 2026. Claude Opus 5.5 previously drew coverage for matching Fable 5.1-class output at a lower cost, a story covered in tech-insider.org’s breakdown of that release’s cost-to-performance math. GPT-6 Astra, Gemini 3.8, and DeepSeek V4.1 Flash have run their own three-way comparison, detailed in separate reporting on that benchmark spread, and Qwen and GLM-class open models have added further downward pressure on what a “reasonable” API rate looks like.
Grok 4.7 does not lead any single outlet’s aggregate ranking based on the data gathered here, but it consistently lands in second or third place on specific coding and agentic benchmarks while undercutting the leaders on price by a wide margin. That positioning, price leader, capability follower, is a strategy xAI has now run across two consecutive releases, and it appears deliberate rather than incidental given that pricing did not change between Grok 4.6 and Grok 4.7 despite the reported parameter increase.
What Developers Should Actually Do With This Information
The practical takeaway from this release cycle is that benchmark-shopping by a single aggregate score, like the Artificial Analysis Intelligence Index, will point away from Grok 4.7. Benchmark-shopping by task-specific accuracy paired with cost per completed task, the way Kingy AI’s CursorBench comparison framed it, tells a much closer story. Teams evaluating an API switch should run their own workload against Grok 4.7’s $2/$6 pricing before ruling it out on the strength of a single index number, particularly for coding-adjacent and legal-document tasks where the published benchmarks show Grok 4.7 within a few points of, or ahead of, its pricier rivals.
curl https://api.x.ai/v1/chat/completions \
-H "Authorization: Bearer $XAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "grok-4.7",
"messages": [{"role": "user", "content": "Summarize this contract clause."}],
"max_tokens": 512
}'
# Standard tier: $2.00 / 1M input tokens, $6.00 / 1M output tokens
# Long-context tier (200K+ token prompts): $4.00 in / $12.00 out per 1M tokensPredictions: Where the Grok-Claude-GPT Price War Goes Next
Based on the pattern established across Grok 4.6 and Grok 4.7, and the broader pricing trend across OpenAI, Anthropic, and xAI through 2026, a few outcomes look likely over the next several release cycles.
- xAI holds the $2/$6 standard rate through at least one more point release, given it has now done so across two consecutive versions despite a reported parameter increase.
- Anthropic and OpenAI face growing pressure to introduce cheaper flagship-adjacent tiers rather than relying solely on smaller distilled models to compete on price, following the pattern already visible in Claude Opus 5.5’s cost positioning.
- Benchmark coverage increasingly pairs accuracy scores with cost-per-task figures, following Kingy AI’s CursorBench framing, rather than leading with a single aggregate index number.
- Enterprise procurement conversations shift toward workload-specific evaluation rather than leaderboard rank, since no model in this comparison set wins across every published benchmark.
- xAI’s roughly monthly release cadence continues to outpace OpenAI and Anthropic’s cycles, raising the likelihood of a Grok 4.8-class release within weeks rather than months of Grok 4.7.
The Bottom Line on Grok 4.7’s Launch
xAI did exactly what it set out to do with Grok 4.7: ship a materially cheaper option than GPT-5.6 Sol or Claude Fable 5.1 without raising prices from the prior generation. It did not close the aggregate benchmark gap to the leading pack, and outlets covering the release, from officechai.com to AI Weekly to Kingy AI, have been consistent in reporting that gap alongside the price story rather than glossing over it. The result is a model that is genuinely competitive on specific coding, agentic, and legal-document tasks, priced well below the alternatives, but not a straightforward like-for-like replacement for teams whose workloads depend on the capability the Intelligence Index is trying to measure.
FAQ: Grok 4.7 Pricing and Benchmarks
When did xAI launch Grok 4.7?
xAI released Grok 4.7 on September 21, 2026, according to xAI’s own announcement and confirmed by multiple outlets including Kingy AI, LLM Stats, and ai-tldr.dev.
How much does Grok 4.7 cost per API call?
Standard pricing is $2 per million input tokens and $6 per million output tokens for prompts under 200,000 tokens, with cached input at $0.50 per million tokens, per xAI’s SpaceXAI developer documentation from September 2026. Prompts of 200,000 tokens or more are billed at $4 input, $1 cached, and $12 output per million tokens. xAI also offers a faster Grok 4.7 Fast tier at exactly double the standard rate, $4 input and $12 output per million tokens, as reported by ComputingForGeeks, for workloads that prioritize latency over cost.
Is Grok 4.7 cheaper than Claude Fable 5.1 and GPT-5.6 Sol?
Yes. Based on figures compiled by officechai.com, Claude Fable 5.1 is priced at $10 input / $50 output per million tokens and GPT-5.6 Sol at $4 input / $20 output per million tokens, both well above Grok 4.7’s $2/$6 standard rate.
Does Grok 4.7 outperform Claude and GPT-6 on benchmarks?
Not on the Artificial Analysis Intelligence Index, where Grok 4.7 scored 46 versus 53 for both GPT-6 Astra and Claude Fable 5.1, per AI Weekly and officechai.com. It does lead or roughly match rivals on specific benchmarks including DeepSWE v1.1, the Harvey Legal Agent Benchmark, and Terminal-Bench 4.0 against GPT-5.6 Sol specifically.
What changed between Grok 4.6 and Grok 4.7?
Pricing and the 500,000-token context window stayed the same. Benchmark scores moved up across several tests, including a two-point gain on the Intelligence Index and a 111-point gain on AA-Briefcase, according to 247wallst.com. AI Weekly’s coverage cited a roughly 40% increase in parameter count behind the release.
Is Grok 4.7 worth switching to for coding tasks?
On CursorBench 4.0, Kingy AI reported Grok 4.7 at its highest effort setting scored 46.3%, close to Claude Opus 5 max’s 46.6%, at roughly half the cost per task ($6.01 versus $11.95). For coding-heavy workloads sensitive to cost, that gap is small enough to justify testing Grok 4.7 directly against current usage.
What is the Artificial Analysis Intelligence Index?
It is a widely cited aggregate benchmark that blends multiple evaluation categories into a single comparative score for frontier AI models, used across the industry as shorthand for overall model capability, though it does not capture cost-per-task or task-specific performance.
Will xAI release a Grok 4.8 soon?
xAI has shipped roughly one release per month through this cycle, with Grok 4.7 landing about a month after Grok 4.6. That cadence suggests another point release is plausible in the coming weeks, though xAI has not published a confirmed date for a next version as of this writing.
Related Coverage
- Unity Launches Official Plugins for Claude Code, Codex [2026]
- Claude Opus 5.5 Matches Fable 5.1 at 60% Less Cost [2026]
- Qwen3.8-Omni-Flash vs Qwen 3.8 Max vs Claude Fable 5.1 [2026]
- Set Up Claude Fable 5.1 API: 12 Steps, 90 Min [2026]
- ChatGPT vs Claude vs Gemini Memory: 2x Capacity Gap [2026]
