Grok 4.5's Coding Debut: Cheaper Than Fable 5, Just Not Quite as Good

xAI shipped Grok 4.5 on July 8, 2026, priced at $2 per million input tokens and $6 per million output tokens — the small chunks of text a model reads and writes, and the unit every AI provider bills by — with a 500,000-token context window. It's xAI's first model built specifically for coding and agentic work, rather than a general-purpose model that happens to code well.

The pitch is the same one every challenger coding model has made this year: comparable results for a fraction of the price. What's different this time is that the fraction shows up twice, in two different benchmarks that measure two different things, and the two results don't agree with each other.

The cost gap is real and it's large

Artificial Analysis, an independent firm that runs its own agent evaluations rather than republishing vendor claims, tracks a Coding Agent Index — a score built by pairing specific models with specific coding harnesses (the surrounding tool, like Claude Code or Codex, that lets a model edit files and run commands) and averaging their results across three tasks: DeepSWE, Terminal-Bench 2.1, and SWE-Atlas-QnA. On that index, Grok 4.5 running inside xAI's own Grok Build harness scores 76 points — one point behind Fable 5 running in Claude Code, and roughly matching GPT-5.5 running in Codex, according to reporting on Artificial Analysis's numbers.

The cost to reach that score is where the gap opens up. Averaged per resolved task, Grok Build costs $2.49, Codex costs $5.07, and Claude Code costs $11.80 — roughly a third of Fable 5's price for a score one point lower:

That gap tracks a token-efficiency difference underneath it: Grok 4.5 averages roughly 1.9 million tokens per task, against 6.2 million for GPT-5.5 and 7.2 million for Fable 5. Fewer tokens per task, at a lower price per token, compounds into the cost spread above.

Where the raw benchmarks disagree

The Coding Agent Index is an average, and averages hide spread. Broken into its parts, Grok 4.5's story is less flattering. On Terminal-Bench 2.1 it scores 83.3%, essentially tied with GPT-5.5's 83.4% and a point behind Fable 5's 84.3%. On DeepSWE 1.1 the gap widens: Grok 4.5 scores 53%, well behind GPT-5.5's 67% and Fable 5's 70%. And on SWE-Bench Pro — a harder, less-gameable successor to the original SWE-bench that grades a model's fix against a real, previously unseen pull request rather than one that might already be in its training data — Grok 4.5 reaches 64.7%, trailing Anthropic's Opus 4.8 at 69.2% and Fable 5 at 80.4%.

None of that is close to embarrassing for a first-generation coding model. It's also not the "matches the frontier at a discount" story the price alone suggests. On the benchmarks that most directly measure whether a model produces a correct fix, Grok 4.5 is a clear step behind the two models it's priced against — a pattern worth checking against DeepSeek's own similarly discounted coding model, where the independently verified numbers likewise trailed the sticker price's implied promise.

The one benchmark it actually wins — with an asterisk

There's an exception, and it cuts the other way. AutomationBench-AA, a benchmark Artificial Analysis built with Zapier to test whether an agent can complete real SaaS workflows — 657 tasks spread across Gmail, Slack, Salesforce, and 37 other simulated apps — scores differently than a typical coding benchmark: a task scores zero the moment a model violates a guardrail (a rule the task explicitly forbids breaking, like emailing a customer without approval), and only counts partial credit for objectives completed if no guardrail was broken at all. It's designed to penalize a model that gets the job done by cutting corners.

On that test, Grok 4.5 comes out on top: a 51.4% headline score against Fable 5's 48.6% and Opus 4.8's 48.5%, at a fraction of the cost — $0.34 per task versus $1.35 for Fable 5 and $1.46 for Opus 4.8. But Grok 4.5 also violates guardrails more often per task (0.63) than either Opus 4.8 (0.55) or Google's Gemini 3.5 Flash (0.46). It wins the average despite that, by completing more of the tasks where it stays inside the rules — which is a real result, not a wash, but it's not the same claim as "breaks fewer rules."

Where this actually fits

Grok 4.5 is available now through Grok Build, Cursor, and the xAI console, with a European rollout targeted for mid-July. For high-volume, lower-stakes coding work — the kind where a wrong answer costs a re-run rather than a production incident — the price and the near-tied Terminal-Bench score make it a reasonable default to try first. For anything closer to unsupervised agentic work with real side effects — sending emails, filing tickets, touching customer data — the AutomationBench-AA numbers say the opposite of what the cost alone implies: cheaper, and not obviously more careless on average, but not obviously more careful per mistake either. Test the specific workflow before trusting the composite score to answer that question for you.