Sofia Lindström
October 12, 2026
12 min read
SpaceXAI, the Elon Musk-led AI unit formerly known as xAI, shipped Grok 4.7 on Sunday, October 11, 2026, pushing its flagship coding model into a frontier race that now includes four separate Claude 5.5-generation releases from Anthropic and an open-weight contender from Mistral. The new model nearly doubled its own predecessor’s score on a long-running coding benchmark, while holding its price steady at the lower end of the market. The release lands in the middle of the busiest six-week stretch for large language model shipping that 2026 has produced, and it says as much about how AI labs are now competing on endurance and cost as it does about raw intelligence.
Don’t miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
SpaceXAI Ships Grok 4.7 as the AI Benchmark Race Accelerates
According to a report aggregated by Ground News, SpaceXAI priced Grok 4.7 at $2 per million input tokens and $6 per million output tokens, with a 500,000-token context window. That is unchanged from Grok 4.6’s pricing structure, which suggests SpaceXAI is competing on capability gains rather than undercutting rivals on cost this round. The company paired its base model with extended reinforcement learning runs deliberately weighted toward harder, longer-duration tasks, including problems described as taking many hours to complete, with the stated goal of improving self-verification and long-context management.
The timing is notable. Grok 4.7 arrived less than two weeks after Anthropic closed out a run of three Claude 5.5-family releases, and just days after Mistral unveiled its open-weight Large 4 model. For developers choosing a default coding assistant, October 2026 has turned into the most crowded release window of the year, with SpaceXAI, Anthropic, OpenAI, Google, and Mistral all shipping within a roughly six-week span.
What Changed: Pricing, Context Window and Coding Benchmarks
The headline number out of the Grok 4.7 launch is a coding-benchmark jump. On Terminal-Bench 4.0, a benchmark that scores how well a model handles software-engineering and terminal-based tasks, Grok 4.7 scored 38.0%, up from 20.3% for Grok 4.6, which is roughly an 87% relative improvement and close to the “nearly doubled” framing used in early coverage. On CursorBench 4.0, a separate benchmark focused on IDE-integrated coding workflows, Grok 4.7 improved from 40.4% to 46.3%. The model also posted a score of 1,657 on Briefcase, an agentic task-completion benchmark, though SpaceXAI did not publish a direct prior-generation comparison for that metric.
Despite the gains, Grok 4.7 still trails the current leader on Terminal-Bench 4.0. Anthropic’s Claude Fable 5.1, released September 1, 2026, scores 57.9% on the same benchmark, meaning SpaceXAI closed part of the gap without actually taking the top spot. Even that leading score comes with an asterisk: benchmark data cited in the Ground News report notes that even the top-performing models still fail more than 60% of the time when tested against private, real-world codebases rather than public benchmark suites. That gap between benchmark performance and production reliability is becoming one of the defining storylines of the 2026 model cycle, and it is a theme Tech Insider has tracked closely in its coverage of Claude Code, Codex, and GitHub Copilot adoption.
Terminal-Bench 4.0 and CursorBench 4.0: The Numbers Behind the Upgrade
Terminal-Bench and CursorBench have become two of the most-cited coding benchmarks in 2026 precisely because they attempt to simulate multi-step, agentic work rather than single-shot code generation. A model that can write one correct function in isolation does not necessarily hold up across a multi-hour task that requires it to run tests, interpret failures, and adjust its own plan. SpaceXAI’s stated engineering focus for Grok 4.7, training the base model against harder and longer tasks through reinforcement learning, is a direct response to that distinction.
The practical upshot for engineering teams evaluating these tools is that benchmark score alone no longer tells the full story. A model can gain significant ground on Terminal-Bench, as Grok 4.7 did, while still sitting well below Claude Fable 5.1 in absolute terms. Teams building around GPT-6.1 Sol, Gemini 4 Argon, or any of the current frontier models are increasingly running their own internal evaluation suites against proprietary codebases rather than relying on published leaderboard numbers alone.
Why SpaceXAI Targeted Long-Running Coding Tasks
SpaceXAI’s decision to weight its reinforcement learning runs toward multi-hour tasks reflects a broader industry shift toward agentic coding, where a model is expected to operate with minimal supervision across an extended session rather than respond to a single prompt. The company also trained Grok 4.7 to natively understand its own “Grok Bot” harness, the scaffolding that lets the model call tools, run code, and manage its own execution environment. That design choice mirrors a move OpenAI made with its Codex harness, which the company opened up as part of its Agents API to support managed, long-running agent services rather than one-off completions.
That convergence, every major lab building bespoke agent harnesses and then exposing them through managed APIs, suggests the coding-assistant market is consolidating around a shared architecture even as the underlying models stay fiercely competitive. For enterprise buyers, it also means harness compatibility and tool-calling reliability are becoming as important a purchasing factor as raw benchmark scores, a dynamic Tech Insider previously examined when covering how Reflection AI’s Beam model approached a similar design tradeoff with far less compute.
How Grok 4.7 Stacks Up Against Anthropic’s Claude 5.5 Family
Anthropic’s release cadence over the six weeks leading into Grok 4.7 was unusually dense. The company shipped Claude Opus 5.5 on September 22, 2026, followed by Claude Fable 5.1 on September 1, 2026 (released ahead of Opus 5.5 in Anthropic’s staggered rollout), then Claude Sonnet 5.5 on September 28, 2026, and finally Claude Haiku 5.5 on October 7, 2026, just four days before Grok 4.7 arrived. On Terminal-Bench 4.0, Claude Sonnet 5.5 scored 70.6%, actually exceeding Claude Opus 5.5’s own reported 66.4% on the same test, an unusual result given Opus is typically positioned as Anthropic’s most capable tier. That inversion has fueled some discussion in developer circles about whether Anthropic’s naming hierarchy still maps cleanly onto raw coding capability.
Claude Haiku 5.5, Anthropic’s smaller and cheaper model, posted a 72.4% score on OSWorld 2.1, a benchmark for computer-use and graphical-interface tasks, up sharply from Claude Haiku 4.5’s 15.7% on the same test. That jump is one of the larger generation-over-generation improvements reported anywhere in the current release cycle, and it is a big part of why Anthropic cut Haiku’s pricing by 75% at launch, positioning it as a volume play for high-frequency, low-cost automation rather than a flagship reasoning model.
Coding Benchmark Comparison Across Current Frontier Models
The table above draws exclusively from figures reported alongside each model’s launch; several cells are marked “not published” because the respective labs did not release a comparable score on that specific benchmark at launch. Even with gaps, the pattern is clear: Claude Sonnet 5.5 currently holds the highest publicly reported Terminal-Bench 4.0 score among models compared directly against Grok 4.7’s own launch data, with Claude Fable 5.1 close behind.
Pricing War: Comparing Token Costs Across the Frontier
Pricing has become as competitive a battleground as benchmark scores. Grok 4.7 holds steady at $2 per million input tokens and $6 per million output tokens with a 500,000-token context window. Claude Sonnet 5.5 is priced at $2 per million input tokens and $10 per million output tokens, matching Grok 4.7 on input cost but running higher on output. Claude Haiku 5.5, by contrast, undercuts both at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, backed by a reported 1-million-token context window, a combination that makes it a clear choice for high-volume, lower-complexity workloads rather than head-to-head coding competitions with Grok 4.7 or Sonnet 5.5.
Mistral’s newly announced Large 4 model, nicknamed “Le Chonk” internally, takes a different approach entirely: it ships with openly available weights rather than a closed API, which changes the cost calculation for teams willing to self-host. Tech Insider covered the model’s 1.05 trillion-parameter architecture and per-token API pricing in detail following its early-October announcement.
The Claude 5.5 Family: Haiku, Sonnet, Opus and Fable Explained
Anthropic’s four-model rollout across September and early October effectively refreshed its entire lineup within six weeks, an unusually fast cadence by the company’s own historical standards. Claude Opus 5.5 remains the company’s largest general-purpose model, while Claude Sonnet 5.5 is positioned as the mid-tier workhorse that, per the reported Terminal-Bench 4.0 scores, currently outperforms Opus 5.5 on coding specifically. Claude Haiku 5.5 is the smallest and cheapest of the four, built for reasoning and coding tasks that do not require frontier-level depth. Claude Fable 5.1, the earliest of the four releases, currently holds the highest publicly reported Terminal-Bench 4.0 score of any model referenced in this comparison.
Developers who want hands-on access to the lineup can follow Tech Insider’s step-by-step setup guides for Claude Opus 5.5 and Claude Sonnet 5.5, both of which walk through API authentication, pricing tiers, and rate limits.
Mistral Large 4 and the Open-Weight Challenge
Mistral’s Large 4 announcement, covered by CNN on October 7, 2026, added a distinct wrinkle to the Grok 4.7 story: it is one of the only frontier-scale models shipping with openly available parameters and architecture, in contrast to the closed-weight approach taken by SpaceXAI, Anthropic, OpenAI, and Google. That distinction matters for regulated industries, defense contractors, and any organization that needs to run inference inside its own infrastructure rather than through a third-party API. CNN’s report also framed the release as part of a broader European effort to compete directly with American labs rather than cede the open-weight category to Chinese developers.
The open-weight dynamic is also reshaping how analysts track the AI race more broadly. Tech Insider has previously reported on how open-h country’s labs currently lead the open frontier, a question that Mistral’s move keeps firmly in play heading into 2027
OpenAI’s Codex Harness and the Rise of Managed Agent APIs
SpaceXAI’s choice to train Grok 4.7 around its own agent harness rather than treat the harness as an afterthought echoes a move OpenAI made earlier in 2026, when it opened its Codex harness as part of a formal Agents API aimed at managed, long-running agent services. That shift signals that labs no longer see the model itself as the full product. Instead, the harness, the tool-calling scaffolding, sandboxing, and session management around the model, has become a competitive differentiator in its own right. Enterprises evaluating coding assistants are increasingly asking not just “how good is the model” but “how well does it integrate with our existing agent infrastructure,” a question Tech Insider explored directly in its look at how DeepSeek V4.1 Flash, Claude Haiku 5.5, and Gemini 3.8 Flash compare on cost-efficient agent deployments.
Historical Context: From Grok 3 to Grok 4.7 in Under Two Years
SpaceXAI’s model line has moved through multiple major and minor version bumps since Grok 3’s debut, with Grok 4, Grok 4.6, and now Grok 4.7 arriving in increasingly tight succession. That pace mirrors the broader industry pattern of 2026, where point releases (4.6 to 4.7, Sonnet 5 to Sonnet 5.5) have replaced the year-long gaps between major versions that characterized the 2023 to 2024 period. The practical effect is that benchmark leadership has become a moving target measured in weeks rather than quarters, with SpaceXAI, Anthropic, OpenAI, and Google each leapfrogging one another on specific tasks rather than any single lab holding a durable lead across the board.
This compressed release cadence also explains why direct comparisons age quickly. A benchmark table built in September 2026 comparing Grok 4.6 against Claude Sonnet 5 would already be two generations out of date on both sides by the time Grok 4.7 and Claude Sonnet 5.5 shipped in October.
Market Impact: What the Release Race Means for Developers and Enterprises
For engineering teams, the practical impact of Grok 4.7’s launch is less about switching providers overnight and more about re-running internal evaluations. Terminal-Bench and CursorBench scores shifted meaningfully across nearly every major model in a six-week window, which means any procurement decision made in August 2026 is likely already stale. Teams relying on Claude Code, Codex, or GitHub Copilot as their primary coding interface will need to weigh whether the underlying model swap is worth a migration, especially given that pricing remained largely stable across this release wave even as capability moved.
There is also a reliability gap worth flagging for any team planning to lean more heavily on agentic coding tools: the reported finding that even top-performing models fail more than 60% of the time on private codebases is a meaningful caveat against over-indexing on public leaderboard scores when making deployment decisions. Benchmark gains are real, but they describe performance on standardized test suites, not necessarily on the idiosyncratic, often poorly documented codebases most enterprises actually run.
Competitive Landscape: SpaceXAI, Anthropic, OpenAI and Mistral Compared
Four distinct strategies are visible in this release wave. SpaceXAI is competing on endurance, training Grok 4.7 specifically for multi-hour agentic tasks while holding pricing flat. Anthropic is competing on breadth, shipping four distinct Claude 5.5-generation models across six weeks to cover every price and capability tier simultaneously. Mistral is competing on openness, betting that open-weight distribution wins deployments in regulated and security-sensitive environments where closed APIs are a non-starter. OpenAI, meanwhile, is competing on infrastructure, turning its Codex harness into a managed Agents API rather than just shipping a better model.
None of the four approaches has produced a single, undisputed leader. Claude Sonnet 5.5’s Terminal-Bench 4.0 lead over its own sibling Opus 5.5 underscores just how fragmented “best model” has become as a question, since the answer now depends heavily on which specific benchmark, task type, and price tier a buyer is optimizing for.
Predictions: Where the Model Race Heads Next
- Expect SpaceXAI to ship a Grok 4.8 or Grok 5 point release within weeks rather than months, continuing the compressed cadence that produced Grok 4.6 and 4.7 in quick succession.
- Anthropic is likely to narrow its own internal Terminal-Bench gap between Sonnet 5.5 and Opus 5.5, since a mid-tier model outscoring the flagship on a flagship benchmark creates pressure to realign the naming or capability hierarchy.
- Open-weight releases, following Mistral Large 4’s lead, will likely keep gaining share among regulated enterprises and government buyers that cannot route sensitive code through a third-party API.
- Agent-harness compatibility, not just raw benchmark score, will increasingly factor into enterprise purchasing decisions as more labs follow OpenAI and SpaceXAI in exposing their harnesses as managed services.
- The gap between public benchmark scores and real-world private-codebase performance will remain a persistent point of scrutiny, pushing more enterprises toward building their own internal evaluation suites rather than trusting vendor-published leaderboards alone.
What Comes Next for Grok and the Broader Model Race
SpaceXAI has not announced a timeline for its next release, and Anthropic has not indicated whether its unusually fast four-model rollout across September and October was a one-time catch-up cycle or the new normal pace for the Claude line. What is clear is that the competitive dynamic has shifted from occasional, headline-grabbing model drops to a near-continuous stream of point releases, each nudging specific benchmarks without any one lab holding a clean, durable lead. For buyers, that means the safest long-term strategy is treating model selection as an ongoing evaluation process rather than a one-time decision, re-testing against internal, private codebases every time a new point release lands rather than relying on launch-day benchmark tables, including the ones published here.
Frequently Asked Questions
What is Grok 4.7?
Grok 4.7 is a coding-focused large language model released by SpaceXAI (Elon Musk’s AI unit, formerly xAI) on October 11, 2026. It is priced at $2 per million input tokens and $6 per million output tokens with a 500,000-token context window.
How much better is Grok 4.7 than Grok 4.6?
On Terminal-Bench 4.0, Grok 4.7 scored 38.0% compared to 20.3% for Grok 4.6, an improvement of roughly 87%. On CursorBench 4.0, it improved from 40.4% to 46.3%.
Does Grok 4.7 beat Claude’s coding models?
Not on the benchmarks reported at launch. Claude Fable 5.1 scores 57.9% on Terminal-Bench 4.0 and Claude Sonnet 5.5 scores 70.6% on the same benchmark, both ahead of Grok 4.7’s 38.0%.
What is SpaceXAI?
SpaceXAI is the renamed identity of xAI, Elon Musk’s artificial intelligence company, which also develops the Grok model family.
Why do even top AI coding models still fail so often?
According to benchmark data cited alongside the Grok 4.7 launch, even top-performing models fail more than 60% of the time when tested against private, real-world codebases rather than public benchmark suites, highlighting a gap between standardized test performance and production reliability.
Is Mistral Large 4 openchitecture, distinguishing it from the closed-weight approach used by SpaceXAI’s Grok, Anthropic’s Claude, OpenAI’s GPT-6 family, and Google’s Gemini models
Which Claude 5.5 model is best for coding?
Based on reported Terminal-Bench 4.0 scores, Claude Sonnet 5.5 (70.6%) currently outperforms Claude Opus 5.5 (66.4%) on coding tasks specifically, even though Opus is positioned as Anthropic’s larger, more capable model overall.
How much does Claude Haiku 5.5 cost compared to Grok 4.7?
Claude Haiku 5.5 is priced at $0.10 per million input tokens and $0.50 per million output tokens, substantially cheaper than Grok 4.7’s $2 per million input and $6 per million output token pricing, though the two target different use cases.
