Grok 4.7 Is Cheap. Its Real Cost Depends on How Many Steps It Takes

SpaceXAI has launched Grok 4.7 with a pitch designed for teams that run AI agents all day: better coding and knowledge-work performance at the same $2 per million input tokens and $6 per million output tokens as Grok 4.6.

The headline sounds simple. The economics are not. A model can have a low sticker price and still produce an expensive result if it needs more tokens, more tool calls or more retries to finish the work. Grok 4.7’s release therefore turns the frontier-model contest into a cost-per-completed-job test rather than a price-per-token comparison.

According to SpaceXAI’s launch report, the new model uses a larger base model, a longer reinforcement-learning run and a harder training mix weighted toward tasks that take hours. It is meant to verify its own work more carefully, manage longer context and operate naturally inside the Grok Bot harness.

The benchmark table shows a meaningful upgrade over Grok 4.6. It also shows why “fastest” or “best” would be the wrong headline. Grok 4.7 wins some specialist tests, lands near the frontier on others and loses clearly on several raw-capability measures.

The $2/$6 price is the real product

SpaceXAI charges $2 per million input tokens and $6 per million output tokens for standard Grok 4.7. GPT-5.6 Sol is listed in the company’s comparison at $4 and $20, while Fable 5.1 is listed at $10 and $50.

Those gaps are large enough to change procurement decisions. Output-heavy agents often generate far more tokens than a chat assistant because they plan, inspect files, call tools, read results and revise their work. At list price, one million Grok output tokens cost 70% less than Sol and 88% less than Fable.

Grok 4.7 Fast doubles output speed and doubles the token price. SpaceXAI’s developer documentation says the fast variant is limited to Cursor and Grok Build rather than the public API. The standard model exposes a 500,000-token context window, text and image input, four reasoning levels and both Responses and Chat Completions APIs.

This is the same strategic lane TECHi identified when Grok Build made xAI a coding-agent contender: xAI is selling a model and an execution environment together. Grok 4.7 makes the bundle tighter.

Grok 4.7 improves sharply, without sweeping the table

On CursorBench 4.0, Grok 4.7 scores 46.3%, up from Grok 4.6’s 40.4% and above GPT-5.6 Sol’s 41.7%. Fable 5.1 remains ahead at 51.8%.

The pattern repeats on long terminal work. Grok 4.7 reaches 38.0% on SpaceXAI’s Terminal-Bench 4.0 setup, nearly doubling Grok 4.6’s 20.3% and narrowly beating Sol’s 37.3%. Fable is much higher at 57.9%.

DeepSWE is closer: Grok 4.7 records 71.0% at high effort, compared with 72.7% for Sol and 70.0% for Fable. The launch materials use xhigh effort for most Grok 4.7 comparisons, with a separate high-effort footnote for DeepSWE. Buyers should reproduce those settings because reasoning effort affects both latency and token use.

The specialist results are more distinctive. Grok 4.7 posts 64.0% on EEBench, ahead of every comparator in the table, and 19.6% on the Harvey Legal Agent Benchmark versus 6.7% for Fable and 2.5% for Sol. It trails both on HealthBench Professional.

That is a useful product profile, not a universal win. Grok 4.7 looks strongest where organizations can exploit its price, electrical-engineering result and legal-agent result. Teams buying maximum coding capability still have reasons to pay more.

Cheap tokens can hide expensive behavior

Token price measures the cost of text passing through the model. It does not measure the cost of a completed migration, security review or debugging task.

An agent can consume extra output by taking many small actions, repeatedly reading the same files or revising plans that do not converge. A cheaper model then looks attractive in a pricing table while using enough tokens to erase part of the advantage.

That concern is especially relevant because Grok 4.7 appears designed for short, rapid interactions with its own harness. Security-testing company XBOW said its early-access results depended heavily on orchestration. The model tended to issue shorter shell commands and take more atomic steps than Grok 4.6, behavior that fit Grok Build better than other setups.

XBOW also reported a real reliability gain: a failure mode in which Grok 4.6 reasoned indefinitely without acting appeared in roughly 0.85% of its runs and in none of the Grok 4.7 runs it evaluated. Removing a low-frequency stall matters when an agent performs thousands of steps.

The two findings belong together. A model that uses smaller actions can recover quickly from feedback and avoid elaborate failed plans. It can also generate more steps. Cost-per-task depends on whether those extra steps reduce retries enough to offset the additional tokens.

SpaceXAI’s launch chart includes tabs for cost, tokens and steps on CursorBench, reinforcing that the company knows the distinction matters. Buyers should insist on all three numbers for their own workloads.

The harness is becoming part of the model

Grok 4.7 was trained to understand the Grok Bot harness natively. That is more than an integration detail. It suggests the model and the agent system are being optimized as one product.

A coding model does not act alone. The harness decides how files enter context, when tools run, how errors return, when history is compressed and whether subagents are available. Changing any of those choices can change the model’s apparent capability and cost.

This makes cross-provider benchmark comparisons harder. A Grok score achieved inside Build may not transfer cleanly to a generic API loop. A competitor’s model may look weaker in an unfamiliar harness and stronger in its own. TECHi’s earlier analysis of Grok on Amazon Bedrock found that availability across platforms does not guarantee identical operational conditions.

The practical buying unit is becoming model plus harness plus pricing policy. Teams should test that unit rather than selecting a model from a leaderboard.

GitHub puts Grok 4.7 in front of more developers

Grok 4.7 is available through the xAI API, Cursor, Grok Build and model routers including OpenRouter, Vercel and Cloudflare. GitHub has also begun a gradual Copilot rollout for Pro, Pro+, Max, Business and Enterprise customers.

GitHub says the model will appear in Visual Studio Code, Visual Studio, Copilot CLI, the Copilot cloud agent, JetBrains, Xcode, Eclipse and the Copilot app. Business and Enterprise administrators can control it through model-policy settings.

That distribution matters as much as a few benchmark points. Developers can compare Grok against other models without changing their editor or rebuilding an agent stack. Usage-based billing also makes the cost difference visible quickly.

The launch revives a question raised during the Grok 3 benchmark dispute: how much confidence should buyers place in vendor-selected evaluations? Grok 4.7’s table is more useful than a single composite score, but it remains a company-published comparison using specific harnesses and effort settings.

A better test for teams considering Grok 4.7

The right evaluation is a basket of real work. Select repeated tasks with known acceptance tests: fixing a production bug, completing a migration, drafting a legal brief from supplied documents, or investigating a security alert.

Measure total input and output tokens, wall-clock time, tool calls, retries, human corrections and the percentage of tasks that pass without intervention. Run the same harness when possible. If the harness must differ, treat it as part of the product being compared.

Grok 4.7 may win that test even when it loses a benchmark. A lower success rate can be acceptable if retries are cheap and fast. A higher benchmark score can be uneconomic if each run consumes several times the output.

The release’s strongest claim is therefore narrower and more credible than “best model.” Grok 4.7 offers frontier-adjacent performance at a price low enough to make sustained agent work practical for more teams.

Its unresolved question is whether the model’s step-heavy style preserves that advantage after the whole workflow is counted. The first buyers to publish cost-per-completed-task data will tell us more than another leaderboard.

What could break the value case

Grok 4.7’s price advantage is most vulnerable in workflows where correctness is hard to verify automatically. A coding agent can run tests and inspect exit codes. A research or legal agent may produce a plausible answer whose error is discovered only during human review. If Grok requires more expert correction than a pricier rival, the labor cost can overwhelm the token savings.

Context use is another pressure point. The 500,000-token window is useful for large repositories and document sets, but filling that window repeatedly is not free. SpaceXAI recommends a prompt cache key so related requests route consistently and can reuse cached input. Teams that omit it may pay full input rates on cache-cold conversations, turning an implementation detail into a recurring bill.

The final risk is platform dependence. Grok 4.7 may perform best in Build because the model and harness were developed together. That can be an advantage for teams willing to adopt the full stack. It is a constraint for buyers that need identical behavior across clouds, regions or internal orchestration systems. A low API price has less strategic value if the best results require an environment a company cannot standardize.

None of those concerns cancels the launch. They define what must be measured during a trial. A useful pilot should report acceptance-test success, reviewer time, cache-hit rate and total cost per accepted result. If Grok wins there, its price-performance claim survives contact with production.

Featured image: Coding photograph by [Yusuf Shamsudeen on Unsplash]( used under the Unsplash License. TECHi crop, color grade, pricing graphic and watermark.