Grok 4.7 vs Claude Opus 5: Data Comparison

Lukas Vogel

Lukas Vogel

Applied Research Editor

Published: September 16, 2026
Comparison of Grok 4.7 and Claude Opus 5 specifications and status

TLDRGrok 4.7 launched on September 21, 2026 at $2/$6 per million tokens; Claude Opus 5 is reported at $5/$25. Compare their benchmark results, pricing, and access.

Opus 5 Ships, Grok 4.7 Cooks: Comparing Grok 4.7 and Claude Opus 5

Grok 4.7 launched on September 21, 2026 as a 2.1 trillion parameter successor to Grok 4.6, while Claude Opus 5 is Anthropic's flagship that shipped July 24, 2026 with a reported Artificial Analysis Intelligence Index score of 51 and third-party pricing of $5 per million input tokens and $25 per million output tokens. Grok 4.7 launched at $2/$6 per million tokens, is available through the Grok API, Cursor, and Grok Build, and has reported strong coding and agent results. Early independent scores still do not make either model a universal winner: Grok 4.7 scored 46 on the Artificial Analysis Intelligence Index, while Opus 5 is reported at 51.

Updated 2026-09-21: Grok 4.7 is now live, with API, Cursor, and Grok Build access reported alongside multiple reasoning variants.

Key Takeaways

  • Both models are live, but their access paths differ. Opus 5 shipped July 24, 2026. Grok 4.7 launched on September 21, 2026 and was reported as available through the Grok API, Cursor, Grok Build, and OpenRouter. Initial launch-day reports said it was not yet available in Grok Chat.
  • The comparison is workload-dependent. Grok 4.7 scored 46 on the Artificial Analysis Intelligence Index, while Opus 5 is reported at 51. Grok's released benchmark results show particularly large gains over Grok 4.6 on coding, terminal work, and legal-agent tasks.
  • The 2.1 trillion parameter figure is now a launch figure. Grok 4.7 is identified in launch reporting as a 2.1T model, compared with Grok 4.6's reported 1.5T. Opus 5's parameter count is not public.
  • Grok 4.7 is cheaper on headline API rates. Grok 4.7 costs $2/$6 per million input/output tokens. Claude Opus 5 is reported at $5/$25 per million input/output tokens. Total task cost can still change with token usage and reasoning settings.
  • Coding evidence is now available for both models. Grok 4.7 is reported at 46.3% on CursorBench 4.0, 38.0% on Terminal-Bench 4.0, and 71.0% on DeepSWE v1.1. Opus 5 has a reported coding score of 78.
  • The verdict is practical rather than absolute. Use the model that fits your workload, toolchain, latency needs, and total token budget. Grok 4.7 is especially compelling for lower-cost coding and agent work, while Opus 5 remains a useful higher-priced reference point.

Grok 4.7 vs Claude Opus 5 at a Glance

DimensionGrok 4.7Claude Opus 5
Public statusLaunched September 21, 2026Released July 24, 2026
DeveloperxAI (SpaceXAI)Anthropic
Model scale2.1 trillion parameters in launch reportingNot yet confirmed
Intelligence Index46 on Artificial Analysis51 on Artificial Analysis, third-party reported
Coding evidenceCursorBench 4.0: 46.3%; Terminal-Bench 4.0: 38.0%; DeepSWE v1.1: 71.0%Coding score of 78, third-party reported
Input price$2 per million tokens$5 per million tokens, third-party reported
Output price$6 per million tokens$25 per million tokens, third-party reported
Context window500,000 tokens reported in Grok Build; formal model-card confirmation remains unclearNot yet confirmed in the supplied record
AvailabilityGrok API, Cursor, Grok Build, and OpenRouter reported; initially not available in Grok ChatPublic/API availability; check terms with Anthropic

This table separates published launch information from community reporting. It treats the released Grok 4.7 benchmark and pricing figures as available launch data, while retaining caveats where access terms or methodology remain unclear.

Availability and Status

Claude Opus 5 has a firm public footprint. It appears in a third-party model index dated July 24, 2026, carrying an Intelligence Index of 51 and a coding score of 78, and multiple community writeups reference it as a shipped, callable model. Grok 4.7 now has the same basic status: it is a released model rather than a pre-launch candidate.

SpaceXAI announced the launch on September 21, describing Grok 4.7 as a notable improvement over Grok 4.6 at the same price and speed. Launch-day reports placed it in the authenticated Grok API, Cursor, and Grok Build. OpenRouter also reported that the model was live. Other integration signals included Hermes Agent and Vercel-related products, although product-specific promotions and integrations should not be confused with universal access.

The rollout was not simultaneous everywhere. One launch-day report said Grok 4.7 was visible in Grok Build but not yet in Grok Chat, while other reports showed the model selector and API access becoming available. The practical distinction is important: "launched" now means the model is publicly rolling out through documented or reported developer and builder channels, not that every Grok product, plan, or region necessarily exposes it at once.

The model is also not open-weight. It cannot be run locally based on the available launch reporting, so teams evaluating it should plan around hosted access and the limits of each integration.

It seems that Grok 4.7 is already being stealth tested in Grok Bot (@bot)! Release is coming closer.

Source: @mark_k

The earlier dated build string, grok-4-7-0907, and apparent stealth testing were pre-launch signals. They are now historical evidence of the rollout process, not the current status of the model. Our earlier breakdown of what the 2.1T signal means for builders should be read in that context.

The Release Delay and Why It Matters

Grok 4.7's launch followed several missed targets. On September 2, Musk said the model would arrive "in 10 days," pointing to around September 12. That deadline passed without a release. Instead, per posts attributed to Musk, the model "needs a few more days to cook."

The stated reason was a training issue, which observers labeled the RL Length Penalty. Reinforcement learning may have penalized response length too heavily, so the model "gives up on hard tasks it can do" too early and does not check its work enough. Harshith framed this as transparency; Teortaxes called the admission "a good one."

The released model's positioning addresses that problem directly. SpaceXAI's launch materials describe longer reinforcement-learning training on harder, multi-hour tasks, more persistence on difficult work, and more careful self-checking. That does not prove that every early-abandonment problem has been solved, but it explains why the delay is part of the model's release story rather than its current availability status.

Capabilities and Model Design

Grok 4.7 is a larger successor to Grok 4.6, built around a new, larger base model rather than only additional post-training on 4.6. Launch reporting identifies it as a 2.1 trillion parameter model compared with Grok 4.6's reported 1.5 trillion parameters. The model is also described as receiving longer reinforcement-learning training on difficult tasks and additional SpaceX engineering data.

The original Token-Efficiency Tradeoff remains part of the comparison, but the launch evidence makes the tradeoff more complicated than the pre-launch pitch suggested. SpaceXAI describes Grok 4.7 as running at the same headline speed and price as Grok 4.6, while some hands-on reports found that high-reasoning runs use substantially more tokens. At xhigh, one report estimated roughly 59,000 reasoning tokens per Artificial Analysis task for Grok 4.7 versus about 20,000 for Grok 4.6. That is a usage observation, not a universal efficiency measurement.

The released feature set emphasizes coding, long-running agents, knowledge work, document and presentation creation, long-context handling, and native understanding of the Grok Bot harness. SpaceXAI also described a rebuilt safeguard system with stronger jailbreak resistance and fewer unnecessary refusals.

The clearest capability signal comes from Musk's earlier roadmap. As Mark Kretschmann relayed, "Grok 4.7 should roughly match Opus 5.0, with multimodal performance still needing work." That forecast was directionally useful, but the released data is more specific: Grok 4.7 is strong on several coding, legal, and engineering evaluations, while its overall Artificial Analysis score of 46 remains below the reported Opus 5 score of 51.

Claude Opus 5, by contrast, is a documented Anthropic flagship that sits below the newer Mythos-tier Fable line but above the Sonnet track. Teams evaluating chat and agent models can call comparable documented options today, including Claude Opus 5 as a reference point for the tier Grok 4.7 is aiming at.

Benchmarks

Grok 4.7 now has a public benchmark record, although the results come from different sources and should not be collapsed into one universal ranking. The reported Artificial Analysis Intelligence Index score is 46, compared with 44 for Grok 4.6. The same score places Grok 4.7 ahead of several models in the cited comparison, while Fable 5.1, GPT-6 Astra, and Opus 5 rank above it.

Reported Grok 4.7 results include:

  • CursorBench 4.0: 46.3%, up from 40.4% for Grok 4.6.
  • Terminal-Bench 4.0: 38.0%, up from 20.3% for Grok 4.6.
  • DeepSWE v1.1: 71.0%, up from 65.2% for Grok 4.6.
  • SWE-Marathon: 46.0%, up from 31.9% for Grok 4.6.
  • HealthBench Pro: 56.7%, up from 48.5% for Grok 4.6.
  • Harvey Legal Agent Benchmark: 19.6%, up from 15.8% for Grok 4.6.
  • EEBench: 66.0%, up from 60.0% for Grok 4.6.

The benchmark profile is therefore uneven but informative. The largest reported gain is on long-running terminal work, where the score nearly doubled. Legal-agent work also shows a substantial increase. CursorBench places Grok 4.7 close to the reported Fable 5.1 result while costing less per task in the cited comparison.

Opus 5 has reported target-model numbers of 51 on the Artificial Analysis Intelligence Index and 78 on a coding score. Those figures are third-party reported and not independently reproduced here. There is still no comprehensive, independently controlled head-to-head evaluation using the same prompts, harnesses, reasoning budgets, and cost accounting for both models.

The Terminal-Bench figure also needs care. The 38.0% result appears in launch and model-card reporting, while a separate discussion said Artificial Analysis's own rerun used a different harness and produced a substantially different result in one setup. That is exactly why benchmark methodology matters more now that both models are live.

Pricing

Claude Opus 5 is reported at $5 per million input tokens and $25 per million output tokens by third-party launch coverage. That is a concrete, comparable tariff.

Grok 4.7 launched at $2 per million input tokens and $6 per million output tokens, the same headline rates reported for Grok 4.6. On token price alone, Grok is less expensive than Opus 5 on both input and output. That difference is material for API users running long coding or agent workflows.

Grok 4.7 also has a fast variant. Launch reporting described the fast mode as delivering approximately twice the output speed at twice the price, although exact access limits, plan entitlements, and total-cost behavior remain unclear. Reasoning variants including low, medium, high, and xhigh were also reported in the API and Hermes-related rollout signals.

The headline tariff does not tell the entire cost story. Several early testers reported heavier reasoning-token consumption from Grok 4.7, so teams should measure cost per completed task rather than compare only dollars per million tokens. A cheaper token can still produce a similar or higher bill if the model uses substantially more tokens.

Which One Should You Use?

Choose Claude Opus 5 if:

  • Your workload has already been validated on its toolchain and you value a reported coding score of 78.
  • You need a higher-priced reference model with an established public footprint.
  • Your tasks favor the behavior and integrations available through Anthropic's API.

Choose Grok 4.7 if:

  • Lower headline API rates matter and $2/$6 per million input/output tokens fit your budget.
  • Your work involves coding, terminal tasks, long-running agents, or engineering-oriented knowledge work.
  • You want access through Cursor, Grok Build, the Grok API, or OpenRouter.
  • You are willing to measure token consumption, reasoning modes, latency, and real task completion rather than relying on one benchmark.

Neither model is categorically better for every workload. Grok 4.7's released results show meaningful gains over Grok 4.6, especially on terminal and agent benchmarks, while Opus 5 remains ahead on the reported overall intelligence and coding scores used in this comparison.

Frequently Asked Questions

Is Grok 4.7 better than Claude Opus 5?

There is no comprehensive independent head-to-head test. Grok 4.7 has a 46 Artificial Analysis Intelligence Index score and strong reported coding and agent results, while Claude Opus 5 is reported at 51 on the same index with a coding score of 78. Grok 4.7 is competitive on some tasks, but neither model is categorically better across every workload.

Is Grok 4.7 cheaper than Claude Opus 5?

Yes, based on the reported API rates. Grok 4.7 costs $2 per million input tokens and $6 per million output tokens. Claude Opus 5 is reported at $5 per million input tokens and $25 per million output tokens. Actual total task cost can still vary with token usage and reasoning settings.

Has Grok 4.7 been released?

Yes. SpaceXAI launched Grok 4.7 on September 21, 2026. It was reported as available through the Grok API, Cursor, and Grok Build, with access also reported through OpenRouter. Initial launch-day reports said it was not yet available in Grok Chat.

How many parameters does Grok 4.7 have versus Claude Opus 5?

Grok 4.7 is identified in launch reporting as a 2.1 trillion parameter model, compared with Grok 4.6's reported 1.5 trillion. Claude Opus 5's parameter count has not been published.

Which should I use for coding, Grok 4.7 or Claude Opus 5?

Both are available options. Grok 4.7's reported results include 46.3% on CursorBench 4.0, 38.0% on Terminal-Bench 4.0, and 71.0% on DeepSWE v1.1, while Claude Opus 5 is reported at a coding score of 78. Choose based on your own workload, tool integration, latency, and total token cost rather than on one score.

When did Grok 4.7 come out?

Grok 4.7 launched on September 21, 2026, after earlier September release targets were delayed.

What to Watch Next

The central release question is resolved: Grok 4.7 is live. The remaining questions are about how consistently the launch claims hold up in production.

First, independent benchmark reproduction matters. The model-card and launch figures show strong gains in several areas, but the Terminal-Bench discussion demonstrates that harness choice can materially change a result. Second, teams should measure total task cost. Grok 4.7's $2/$6 tariff is substantially below Opus 5's reported $5/$25 rates, but heavier reasoning-token use can narrow the practical advantage. Third, access details still need clarification across Grok Chat, subscriptions, API limits, Cursor plans, Grok Build, and fast or xhigh variants.

The context figure also deserves a formal answer. A 500,000-token context window was reported in Grok Build, but the supplied record does not establish a complete official context-window specification or whether the same limit applies across every access path.

Update — 2026-09-20

Since publication, multiple observers reported a base_model:grok-4.7 entry in Google Cloud’s quotas and system-limits pages. That is a stronger deployment-preparation signal than the earlier bot sightings, but it still does not establish a public API or consumer rollout. HarshithLucky3 reported the quota entry and LuminaBench documented the Google Cloud listing.

Testing signals also moved beyond the earlier reports: LuminaBench said a benchmark site briefly exposed grok-4.7-build at xhigh before removing it under an embargo, while NFT_Chen circulated purported KernelBench results showing uneven comparisons across kernels. Neither item is an official benchmark or model card, and repeated September 17 launch predictions still did not produce a confirmed public release.

Update — 2026-09-21

On September 21, Kimmonismus reported that Grok 4.7 had appeared in OpenCode Zen and was subsequently deleted. A separate report also placed it briefly in the OpenCode Go catalogue. These transient third-party listings strengthen the signal of staged access, but they still do not establish a public xAI release, API availability, or official model documentation.

New infrastructure evidence extends beyond the earlier catalogue sightings. LuminaBench reported that an authenticated Hermes pull exposed grok-4.7 with low, medium, high, and xhigh reasoning variants. Separately, LuminaBench said a Cursor Agent pull request explicitly referenced both Grok 4.6 and Grok 4.7 and said the newer model had been benchmarked on multiple sites.

TokenGremlin’s aggregation adds a purported grok-4-7-0907-xhigh identifier in Cursor logs to the previously reported backend and quota traces. These are stronger signs of staged or internal access across several services, but they remain third-party infrastructure reports: they do not establish a public xAI launch, production API, official pricing, or published benchmark results.

Try related models on Kie.ai: Grok 4.7, Claude Opus 5.5 and GPT-6 Sol and Luna.

Lukas Vogel

About Lukas Vogel

Lukas reads the papers and model cards so you do not have to, focusing on reproducible claims.

View all posts by Lukas Vogel