SpaceXAI Releases Grok 4.7 as Model Competition Shifts Toward Long-Horizon Tasks

SpaceXAI has released its latest flagship model, Grok 4.7, targeted at coding, agentic workflows, and specialized knowledge work. Retaining the pricing and output speed of Grok 4.6, the standard tier is now available in Cursor, Grok Build, and the xAI API.

Rather than framing this release around generic benchmark gains, xAI shifted its training focus toward tasks requiring hours to complete: whether a model can continuously invoke tools, verify its own work, and maintain goal stability across extended contexts. Understanding Grok 4.7 begins with this shift.

What Grok 4.7 Aims to Solve: Why Long Tasks Derail Mid-Execution

Short tasks easily create the illusion of model capability: writing a function, explaining a snippet of code, or generating a page of copy often yields results within one or two turns. The real performance gap emerges in long-horizon tasks—where a model must sequentially read files, call tools, resolve errors, retain preceding constraints, and still know what it is solving dozens of steps later.

To address this, Grok 4.7 adopts a larger base model than Grok 4.6, extends reinforcement learning training, and skews the task distribution toward problems requiring hours to complete. xAI summarizes the improvements into two key areas: more meticulous self-verification and better management of long contexts.

It has also been trained to natively understand the Grok Bot harness. A harness can be thought of as the model's external runtime environment: responsible for structuring conversations, invoking tools, executing commands, and persisting task state. This specifically targets Grok Bot's native operational patterns, aiming to make the model more familiar with how tasks are organized and driven forward in dialogue and general knowledge work.

The inference configuration provides four reasoning effort tiers: low, medium, high (default), and xhigh. This allows teams to handle routine tasks with lower computational investment while reserving more compute for complex debugging, long-chain agents, and specialized domain work.

Outperforming 4.6 Across Seven Benchmarks, with the Largest Leap in Terminal Tasks

This table illustrates Grok 4.7's position most clearly: while it does not take first place across every metric, it avoids any trade-offs where gains in one area come at the expense of regression in another compared to the previous generation.

• Long-horizon coding: CursorBench 4.0 improved from 40.4% to 46.3%.
• Software engineering: DeepSWE v1.1 increased from 65.2% to 71.0%.
• Multi-hour terminal work: Terminal-Bench 4.0 rose from 20.3% to 38.0%, an increase of 17.7 percentage points and a relative gain of roughly 87%.
• Electrical engineering: EEBench improved from 53.0% to 64.0%.
• Legal tasks: Harvey Legal Agent rose from 15.8% to 19.6%.
• Clinical reasoning: HealthBench Professional increased from 48.5% to 56.7%.
• Multi-hour office tasks: AA Briefcase v1.1 advanced from 1,546 to 1,657.

Among these, the change in Terminal-Bench aligns most closely with the core training narrative of this release. It does not measure a single code completion, but rather whether a model can sustain multi-step execution within a terminal environment. A score of 38.0% also implies that such tasks are far from "solved," yet compared to 4.6, the probability of the model sustaining productive execution across long chains has improved noticeably.

From a peer comparison perspective, Grok 4.7's advantages center on pricing and specific professional tasks. It leads Fable 5.1 and GPT-5.6 Sol in EEBench and Harvey Legal Agent; Fable 5.1 scores higher on CursorBench, Terminal-Bench, AA Briefcase, and HealthBench, while GPT-5.6 Sol is slightly higher on DeepSWE and HealthBench. When comparing, column labels are also worth noting: Grok 4.7 uses xHigh, 4.6 uses High, and the other two models use Max; on DeepSWE, Grok 4.7 is labeled as High Effort.

The Second Anchor for Long-Horizon Capability: Delivering Complete Documents

Coding and professional documentation may appear to be distinct categories of work, but over extended durations they suffer from the same failure mode: the former must keep files, dependencies, and runtime states strictly consistent, while the latter must maintain cohesion across source materials, figures, arguments, and layout formatting. If a model forgets an earlier constraint, the downstream deliverable degrades as a whole.

This is precisely why xAI positions knowledge work alongside coding. Grok 4.7 is designed not merely to draft locally correct snippets, but to synthesize scattered materials into cohesive, end-to-end documents and presentations.

AA Briefcase tasks models with multi-hour office workflows representative of lawyers, nurses, financial analysts, and other professionals; GDPval similarly evaluates specialized knowledge work. In official charts, Grok 4.7 achieves a GDPval Elo of 1,695, up from 1,605 for Grok 4.6 and approaching Fable 5.1 Max at 1,735. This indicates that the upgrade trajectory extends beyond "writing better code" to ingesting longer source materials, structuring information, and delivering complete end products.

This also explains why xAI emphasizes both self-verification and long-context management. For documentation, analysis, and presentation tasks, the challenge rarely lies in producing an isolated eloquent sentence; it lies in whether the model can maintain standard definitions across extensive materials and carry earlier evidence accurately through to downstream conclusions.

True Pricing Advantage Depends on the Cost per Completed Task

Looking strictly at per-token pricing only answers whether each individual step is expensive, not how much it costs to complete a task end-to-end. A more capable model that takes fewer detours can leverage higher unit rates into lower aggregate costs; conversely, a cheaper model that repeatedly retries can end up more expensive on long-horizon tasks.

The CursorBench 4.0 curves published by xAI include multiple operating points for the same models. While Fable 5.1 reaches higher absolute scores, it does so at higher average task costs; within comparable cost bands, Grok 4.7's curve sits above those of GPT-5.6 Sol and Sonnet 5. This embodies what xAI terms the "price-performance frontier": the primary value proposition is not sweeping every score, but completing more long-horizon programming tasks within a given budget.

For engineering teams in production, the most critical metric to track should shift from per-token rates to the total cost per successful task: factoring in model invocations, retry attempts, tool execution overhead, and human interventions is the only way to determine whether an upgrade genuinely saves money.

Why Open-World Games Are Harder Than Writing a Function

An open-world game is not a single feature; it is an ensemble of systems that must function simultaneously: vehicles must navigate roadways, pedestrians and traffic must share world state, and camera angles, minimaps, and speedometer UIs must react to player input in real time. Individually, none of these components appear overwhelmingly complex; the true difficulty lies in preventing them from conflicting with one another throughout sustained development.

In its release thread, SpaceXAI demonstrated this workload through an urban driving showcase. The vehicles, pedestrians, streets, buildings, minimap, and speedometer UI on screen reflect continuous multi-module coordination rather than a one-off function completion.

This scenario makes the concept of "multi-hour tasks" tangible: the model must not only generate greater volumes of code, but also preserve long-term goals, state, and interdependencies across multiple sub-tasks. The architectural adjustments in Grok 4.7's training are designed specifically to minimize mid-task context drift, conflicts, and cyclical rewrites.

Safety Capabilities Begin Shipping Alongside Task Capabilities

Grok 4.7 deploys an updated safeguard stack. xAI set dual objectives: strengthening refusal rates for hazardous requests while minimizing false refusals against legitimate cybersecurity work.

The company shared two specific findings: a 62.4% refusal rate on the LatchBio biosecurity benchmark, and 3.3% of high-risk dual-use prompts permitted on HackerBench v0.3. xAI has also begun offering invite-only Grok 4.7 red-teaming capabilities to selected cybersecurity partners for defensive research.

The significance of this track is clear: as long-horizon agents gain broader tool permissions, safety can no longer remain a matter of chat-interface content filtering. Because models execute shell commands, manipulate files, and advance long-running workflows, the ability to identify dangerous actions alongside maintaining usability for legitimate operations must both be evaluated as core product metrics.

Key Configurations for Production Deployment

Standard pricing matches Grok 4.6:

• Below 200k prompt tokens: $2 input, $0.50 cached input, and $6 output per 1M tokens.
• Above 200k prompt tokens: $4 input, $1 cached input, and $12 output per 1M tokens.

Grok 4.7 Fast uses the identical model on accelerated infrastructure, delivering 2x standard output speed at 2x token pricing. It is available exclusively within Cursor and Grok Build, does not draw from Grok Build free quotas, and is not exposed via the public xAI API.

Beyond baseline model capabilities, three runtime configurations directly affect long-horizon workflow costs and continuity:

1. Ensuring Cache Hits: Setting prompt_cache_key in the Responses API or x-grok-conv-id in Chat Completions routes requests from the same session to the same server, preventing cache misses that incur full input token charges repeatedly.
2. Managing Context Growth: For deep tool invocation loops and prolonged iterations, pairing runs with context compaction compresses conversational history, reducing downstream latency and inference cost.
3. Persisting Multi-Turn Reasoning State: The Responses API returns reasoning.encrypted_content; passing these reasoning items back unmodified in subsequent calls allows the model to retain and build upon that state across multi-turn tasks.

Teams requiring US data residency can also target https://us.api.x.ai/v1, ensuring inference remains within the United States for a 10% premium on token pricing.

How Teams Should Evaluate Adoption

The most revealing way to evaluate Grok 4.7 is not through single-turn Q&A or isolated code completions, but on tasks that currently derail mid-execution: multi-file refactoring, persistent terminal debugging, extensive source document analysis, and comprehensive report or presentation synthesis.

Teams can evaluate their current models against Grok 4.7 on the same set of real-world workflows, tracking three key metrics:

1. End-to-End Success Rate: Whether tasks genuinely complete against defined acceptance criteria.
2. Total Cost per Successful Task: Combined expenses across model tokens, retries, tool execution cycles, and human labor hours.
3. Human Takeover Point: The exact step at which the model begins dropping constraints, repeating errors, or forcing a human engineer to intervene and replan.

Only when success rates rise, total costs decline, and human takeovers decrease does Grok 4.7's long-horizon training translate into tangible engineering value. The shift it represents is not just "another smarter chat model," but the progression of model evaluation toward a more demanding standard: who can reliably finish complex work at lower total expense.