FULL STORY

Grok 4.7: From Leak to Launch and Benchmarks

Grok 4.7 briefly surfaced in OpenCode Zen's directory before xAI officially launched it on Sept 22 with same price and speed but big benchmark gains. Early hands-on praise was tempered by token-usage and independent benchmark controversies.

2026-09-21 ~ 2026-09-23 · 9 episodes · 136 posts

Episode 1 · Grok 4.7 Briefly Surfaces in OpenCode Zen Directory, Fueling Launch Rumors (2026-09-21, 5 posts)

On September 21, signs pointed to an imminent launch of xAI's next-generation model Grok 4.7: multiple X users spotted grok-4.7 briefly listed in OpenCode Zen's model directory before it was removed, widely read as a pre-launch signal. The same day, leakers claimed an afternoon release with specific specs and pricing. xAI has made no official confirmation.

Confirmed

  • grok-4.7 did appear in OpenCode Zen's model directory and was subsequently deleted, reported by @kimmonismus, @CurieuxExplorer, and @socialwithaayan; @CurieuxExplorer joked that being on the menu means it's already in the kitchen.
  • xAI has not confirmed a Grok 4.7 launch as of now.

Unconfirmed

  • @ns123abc claimed Grok 4.7 would launch early that afternoon, citing a screenshot, with no official backing.
  • @realsohamparekh leaked that Grok 4.7 launches today with a 500K-token context window, multimodal support, and pricing close to Grok 4.6 — unverified personal-account claims.
  • @kimmonismus and @socialwithaayan both speculated this week is a "major model release week," with OpenAI (Dev Day) and possibly Anthropic to follow — personal speculation.

Why it matters

  • A model name briefly appearing in OpenCode Zen's directory is typically seen as a precursor to a vendor's official launch, so the community treats this as a strong signal.
  • If the 500K context, multimodality, and Grok 4.6-level pricing rumors hold, Grok 4.7 would directly compete on long-context and cost-effectiveness — pending xAI's official announcement.

Episode 2 · xAI launches Grok 4.7 with major benchmark gains at unchanged price and speed (2026-09-21, 59 posts)

xAI officially launched Grok 4.7 on September 22, calling it its strongest coding and knowledge-work model to date, with significant improvements at the same price and speed as Grok 4.6. The model is available in Grok Build, Cursor and other products, and the model card shows major benchmark gains, but third-party early tests aggregated by TheAIGRID (such as ValsAI) suggest a gap with some community expectations; real-world reception remains to be seen.

Confirmed

  • Announcement and availability: a Reddit post relayed xAI's official news page on September 21; the official SpaceXAI account formally announced the release on September 22, with Elon Musk repeatedly amplifying it. Available in Grok Build (Low/Medium/High/Extra High reasoning levels) and Cursor; grok update upgrades the CLI.
  • Pricing unchanged from 4.6: $2 per million input tokens, $6 per million output tokens; the company also claims twice the speed at half the price of comparable models.
  • Improvements: a new, larger base model with longer reinforcement-learning training on multi-hour tasks, better self-verification and long-context retention, native understanding of the Grok Bot system, and what xAI calls its strongest safety guardrails to date. Team member haozhuwang noted training on longer-horizon problems.
  • Benchmarks (model card, shared by Musk): Terminal-Bench 20.3% → 38.0%; SWE-Marathon 31.9% → 46.0%; CursorBench 4.0 40.4% → 46.3%; large gains on Harvey legal benchmarks and HealthBench Pro.
  • Official 4.6 vs 4.7 demos: building an open-world city game and building Age of Empires II with AI.
  • Real-world case: Box's Agent preview used Grok 4.7 on a $2M insurance claim, flagging $82,000 in duplicate invoices.
  • Team and ecosystem: team member aksheyd cited extensive harness polish and partnerships with top engineering teams; early users praised coding, engineering and 3D performance at high TPS in Grok Build and Cursor.
  • NVIDIA congratulated the launch and said it provides accelerated compute for xAI.
  • Early third-party feedback: TheAIGRID aggregated posts noting ValsAI and other testers found overall results below community expectations.

Unconfirmed

  • @AlexFinn's claim that Grok 4.7 is an "agentic model" trained specifically for Grok Bot to do knowledge work comes from a single account and is not in the official announcement (though native Grok Bot support is confirmed).
  • @Polymarket's "twice the speed, half the price" framing conflicts with official pricing parity with 4.6; the official announcement should be treated as authoritative.
  • Pre-launch double-speed-tier pricing reported by @markk was not mentioned in the official announcement.

Why it matters

  • Same price and speed plus near-doubled Terminal-Bench and large CursorBench/legal gains directly boost Grok's cost-performance edge in agent, terminal and professional knowledge-work scenarios.
  • Emphasis on multi-hour tasks, self-checking and native Grok Bot integration, plus Box's claims-audit demo, signals xAI is betting on full-agent workflows for real work.
  • Commenters flagged its electronics-engineering results—SpaceX's engineering data may be a unique training advantage.
  • The gap between official messaging and early third-party tests suggests real-world performance beyond formal benchmarks will decide its competitiveness.

39 more related posts →

Episode 3 · Grok 4.7 Day-One Hands-On: Strong Word-of-Mouth for Autonomous Coding (2026-09-22, 17 posts)

On September 22, xAI released Grok 4.7, drawing heavy day-one hands-on sharing from developers with largely positive reception—most saw it as a major step up from Grok 4.6. Initially it was only available in Grok CLI, with the web version and Grok bot not yet live (vista8).

Confirmed

  • Developer ScriptedAlchemy reported Grok 4.7 working autonomously on a single goal for over 70 hours, with much better attention to detail and effective use of the 500K-token context (m3)
  • Developer Matt Shumer called 4.7 a very capable daily-driver model and a huge improvement over 4.6 (m2)
  • Developer tetsuoai reported Grok 4.7 building a custom game engine and playable FPS in pure C (SDL2 and OpenGL) in about 7 hours, calling it his favorite coding model and recommending Grok Heavy mode with maxed tokens (m4, m5)
  • One developer noted markedly better system-prompt adherence, pushing back on negative coverage citing poor public benchmarks, arguing benchmarks and 3D game demos don't reflect real work capability (m1)
  • Developer DanielFarinax tested procedural watch exploded-view and HTML water-glass generation with solid results; he is also rebuilding his Grok-driven browser Xplor with 4.7, achieving things 4.6 couldn't, though he felt 3D tasks were weaker and still needed extreme prompting (m8, m14)
  • Developer Baconbrix asked for a procedural game map based on Boca Chica, Texas; the model ran overnight and auto-added biomes, day-night cycles, fireflies and swamp-bubble particles, impressing him (m15)
  • vista8 generated a playable Plants-vs-Zombies-style web game in one prompt, with zombies and items drawn in SVG in an "ugly-cute" style (m11, m12, m17)
  • Users compiled 10 day-one extreme use cases, including a Boost Grok Bot productivity workflow, catching a $146,000 error for a business, Blender scenes, cloth simulation and robotic-arm animation, with commenters asking how to access the 500K-token context; CodeByPoonam also rounded up 7 community projects within 24 hours (m7, m10, m13, m16)

Unconfirmed

  • A user posted a comparison video labeled Grok 4.7 vs 4.6 building open-world city games, showing 4.7 clearly stronger in complex 3D modeling, collision physics, kinematics and lighting—dubbed the "most fun benchmark"—but the version label is unverified (m6, m9), and it sits in tension with DanielFarinax's weaker 3D impressions
  • DanielFarinax's account of rebuilding Xplor with 4.7 is self-reported and unverified (m14)

Why it matters

Day-one testing focused on real engineering scenarios—long autonomous work, ultra-long context use, and low-level systems programming—rather than benchmark scores, offering third-party evidence for evaluating frontier-model productivity. If 500K-token context effectiveness and 70-hour autonomous runs are further validated, they would be an important signal about the capability boundaries of AI coding agents.

Episode 4 · Grok 4.7 Launches: Big Benchmark Gains but Doubled Token Use Sparks Cost Debate (2026-09-22, 25 posts)

Grok 4.7 was announced by Elon Musk on 09-22 and immediately put through third-party benchmarks, led by Artificial Analysis (AA). The headline conclusion: substantial capability gains paired with sharply higher token usage. AA scored it 46 on the Intelligence Index (2 points above Grok 4.6), putting xAI in the top four labs for the first time; on the multi-hour office benchmark AA-Briefcase it took second place with 58%, just 1 point behind Claude Fable 5.1 Max (59%) and ahead of GPT-6 Astra. But at roughly 81k output tokens per task—more than double its predecessor—its per-task cost exceeded Astra in AA's comparison, fueling a community debate over value.

Confirmed

  • AA evaluation: Intelligence Index 46 (Grok 4.6 at 44); hallucination rate down from 34% to 29%; output speed 188 tokens/s, 7.1 minutes average per Intelligence Index task
  • AA-Briefcase: 1657 Elo, +111 over Grok 4.6 (high), behind only Claude models; analysis-quality Elo at 1994, +88 over GPT-6 Astra; GDPval-AA ahead of Astra by +153 Elo
  • Win-rate: 58%, second only to Claude Fable 5.1 Max at 59%
  • AA stated that in the AA-Briefcase due-diligence scenario, Grok 4.7's per-task cost is about half of Opus 5's
  • Coding: CursorBench 4.0 at 46.3%, nearly matching Opus 5 Max's 46.6%; CursorBench up from 40.4%, EEBench from 53% to 64%; Coding Agent Index rose from 47 to 56 with the Grok Build agent
  • Cost: 81k output tokens per task for xhigh, over double Grok 4.6 (xhigh); AA's comparison put per-task cost above GPT-6 Astra
  • Showcase cases: independently running a step-by-step valuation chain and challenging the counterparty's estimate; analyzing three full years of financials to expose FX-masked growth, where 4.6 used only the latest year
  • FinanceYF5 compared open-world city game demos by 4.7 vs 4.6, finding 4.7 worked longer, verified more, and produced better output at unchanged pricing

Unconfirmed

  • XFreeze's personal benchmark: on an identical coding task, Grok 4.7 xHigh matched Opus 5 Max at under half the cost with fewer tokens and steps—a personal measurement that doesn't contradict AA's Astra comparison
  • One blogger's same-prompt test at max reasoning: Grok 4.7 took 32 minutes and $8.14 versus free SWE 2 at 15 minutes and $0
  • ChrisGPT claimed Grok 4.8 arrives by end of November and that Grok 4.7's per-task cost exceeds GPT's—unverified leaks

Why it matters

  • Users including Angaisb and Theo (repost) noted Grok 4.7 uses nearly 3x Astra's tokens per task and 2.25x Grok 4.6's, meaning real costs under "same price, same speed" marketing may be higher; Angaisb argued OpenAI is the only lab serious about token efficiency, and Theo's repost said Grok 4.7 is no longer cheap, costing more than Astra in AA's comparison
  • Critics (relayed by ivanbezdomny) questioned the value of a 2-point gain at 2.5x token consumption
  • For heavy API users, the model now matches frontier capabilities at higher unit cost, requiring workload-specific recalculations; meanwhile its deeper multi-hour agent analysis is a clear capability gain for professional work

5 more related posts →

Episode 5 · Grok 4.7 Reportedly Closes Gap With Top Frontier Models at Lower Cost (2026-09-22, 2 posts)

Users claim Grok 4.7 has nearly closed the gap with the most expensive frontier models and even leads in some areas at far lower cost, though xAI has not confirmed this. Commentators caution that capability claims should be weighed against pricing.

Episode 6 · Open-Source Grok 4.7 Editor Extension Passes 140K Installs (2026-09-22, 3 posts)

An unofficial open-source extension bringing Grok 4.7 to VS Code, Cursor, and Antigravity appeared on launch day via Open VSX and the VS Code Marketplace, surpassing 140K installs. Developer Pawel Huryn confirmed the channels while promoting his AFK Pilot project.

Episode 7 · Bug Hunt Bench: GPT-6 Astra Tops 105-Bug Fixing Test, MiMo Stuns on Value (2026-09-22, 18 posts)

Independent evaluator Pawel Huryn (author of The Product Compass) built and published Bug Hunt Bench: 105 bugs drawn from two real code repositories, with frontier models asked to find and fix them, scored by blind judges from different model families against a private answer key. GPT-6 Astra (max) topped the board fixing 45 ($33.03, n=3); Opus 5.5 matched GPT-6 at roughly half the cost; Xiaomi's newly released MiMo-V2.6-Pro scored 22.7 for just $0.86, which Huryn called the value benchmark; Grok 4.7 showed no meaningful gain over 4.6.

Confirmed

  • Method: two real repos, one prompt per repo, blind cross-family judging against an answer key, unfixed/missed bugs scored 0, multiple runs per model with statistics reported, published at Bug Hunt Bench (bughunt.productcompa…).
  • Results: GPT-6 Astra (max) 45, first ($33.03, n=3); Opus 5.5 ties GPT-6 at about half the cost; Muse Spark 1.3 (max) 32.2 ($18.11); GPT-5.6 Luna (max) 31.5 ($2.82); MiMo-V2.6-Pro 22.7 ($0.86); Grok 4.7 and 4.6 both 28.7. Costs differ by nearly 200x.
  • Interim leaderboard (m8, m10): GPT-6 Astra (max) 45 (n=3), Fable 5.1 (max) 43 (n=1), GPT-5.6 (max) 42.5, Opus 5.5 (max) 41.7; GPT-6 Sol (max) fixed only 29.3, a sharp regression to last place (m1, m9); GPT-5.6 Sol (max) 43.5. Huryn also noted Opus 5.5 loses more score at lower reasoning effort.
  • Difficulty clarification (m14): the 105 bugs were not randomly sampled but hard problems frontier models initially failed in early 2026; real bugs not deliberately planted are excluded because OpenAI would flood results with real, theoretical bugs.
  • Grok follow-up (m11, m15): at high reasoning, Grok 4.6 beat 4.7 by 23 to 19, and 4.6's medium tier was also better; at max, a fourth run (n=4) put 4.7 ahead by only 0.1, judged random noise, and testing at that tier stopped; xhigh showed no separation in three runs.

Unconfirmed

  • Huryn's suspicion that Grok 4.7 was optimized for popular benchmarks rather than real coding skill remains personal speculation without corroboration.
  • Whether this two-repo, hard-bug-heavy benchmark generalizes to real engineering awaits independent replication; praise of MiMo's value by reposters like GregGL87 is secondhand commentary.

Why it matters

  • By deliberately selecting problems all frontier models initially failed and publishing costs, the benchmark vividly exposes huge capability and price gaps (MiMo-V2.6-Pro got roughly half of GPT-6's score at 1/38 the price), giving developers a third-party reference distinct from official marketing.
  • Grok 4.7 shows no real generational gain here and GPT-6 Sol regresses sharply versus Astra — valuable independent signals for anyone tracking frontier model progress.

Episode 8 · Grok 4.7 Meets Xiaomi MiMo-V2.6 in a Same-Night Price War (2026-09-22, 5 posts)

In the early hours of September 22, xAI's delayed Grok 4.7 and Xiaomi's MiMo-V2.6 launched one after the other. Third-party evaluations showed similar capabilities, but MiMo-V2.6's output price is roughly 1/16 to 1/20 of Grok 4.7's, and multiple reviewers considered Grok 4.7 below expectations.

Confirmed

  • Five days after livestreaming its RL training on the 19th, Xiaomi released three MiMo-V2.6 variants—Pro, Flash, and Pro-UltraSpeed—with UltraSpeed output up to 20x faster than Pro; Pro costs 0.025 yuan per million cached-hit input tokens. MiMo-V2.6 topped the open-source leaderboard (@卡尔的AI沃茨).
  • Grok 4.7 is a newly pretrained 2.1T-parameter model with a 500K context; Cognition brought it to Devin, scoring 59.4% on FrontierCode 1.1 Extended, which the company says performs well on hard backend engineering tasks (@airesearch12).
  • Pricing: Grok 4.7 output costs up to ¥40/M tokens versus about ¥2 for MiMo-V2.6—roughly 16x cheaper, by some counts 20x (@op7418, @歸藏的AI工具箱, @卡尔的AI沃茨).

Reactions

  • @op7418 observes that model releases now split into two extremes—cheapest or most frontier—or they get no traction.
  • @数字生命卡兹克 tested Grok 4.7 and found it below expectations with an AA index of 46; @airesearch12 relays users calling Grok 4.7 disappointing given MiMo's similar intelligence at far lower cost.

Why it matters

Two models released the same day with near-equal capability but an order-of-magnitude price gap highlight open-source budget models directly challenging closed-source frontier models—"a new open-source leader at half the price" is the core narrative of this round.

Episode 9 · xAI Ships Three Grok Frontier Models in Nine Weeks (2026-09-22, 2 posts)

xAI released three frontier models within roughly nine weeks: Grok 4.5 on July 16, Grok 4.6 on August 12, and Grok 4.7 on September 21. The rapid cadence has drawn widespread attention.