Grok 4.5 Launches: Frontier Coding at Half the Cost

By Radar — independent tracking of frontier AI releases and what they mean for builders.

What is Grok 4.5 and why does it matter?

Grok 4.5 is SpaceXAI's newest frontier model, released on 8 July 2026 and trained alongside the Cursor coding editor on real developer workflows. It leads rivals on the AutomationBench agent test at 51.4% and prices at $2 per million input tokens and $6 per million output, roughly a third to a fifth of Opus 4.8, GPT-5.5 and Fable 5 — making frontier-grade agents genuinely affordable.

TL;DR

• Priced to disrupt: $2/M input, $6/M output (The Decoder, 2026) — versus $5/$25 for Opus 4.8 and $10/$50 for Fable 5.
• #1 on AutomationBench-AA: 51.4%, ahead of Fable 5 (48.6%) and Opus 4.8 (48.5%), at $0.34 per task versus ~$1.40 (Yellow, 2026).
• Token-frugal: uses about 4.2× fewer output tokens than Opus 4.8 on SWE-Bench Pro (MarkTechPost, 2026).
• Trained on Cursor: trillions of tokens of real developer-agent interactions, not just static code (Cursor, 2026).
• Not the raw benchmark king: it trails Fable 5 on pure coding scores — but wins decisively on cost per completed task.

How much does Grok 4.5 cost compared with rivals?

Price is the headline, and it is not close. Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, with cached input at roughly $0.50/M (The Decoder, 2026). Set against the current frontier, that is a step change:

| Model | Input ($/M) | Output ($/M) | |---|---|---| | Grok 4.5 | $2 | $6 | | Opus 4.8 | $5 | $25 | | GPT-5.5 | $5 | $30 | | Fable 5 | $10 | $50 |

Sources: The Decoder, 2026; MarkTechPost, 2026.

But sticker price understates the gap, because Grok 4.5 also spends fewer tokens to finish the same work. On the Coding Agent Index, a single task cost $2.49 in Grok Build, versus $5.07 for GPT-5.5 in Codex and $11.80 for Fable 5 in Claude Code — Grok burned 1.9M tokens per task against 6.2M and 7.2M respectively (The Decoder, 2026). That compounding of a lower rate and lower consumption is where the "half the cost" framing actually understates things for agentic loops.

Is Grok 4.5 really as good as Opus 4.8 at coding?

Honestly? On raw coding benchmarks, not quite — and SpaceXAI doesn't pretend otherwise. On SWE-Bench Pro it resolves 64.7% of tasks, trailing Fable 5's 80.4% (max), and on Terminal-Bench 2.1 it scores 83.3% against Fable 5's 84.3% (MarkTechPost, 2026). Independent analysts peg it at roughly "Opus 4.7 level" — about one generation behind the very top on pure capability.

Where it flips the table is agentic, tool-using work — the messy, multi-step tasks real products actually run. On the AutomationBench-AA benchmark, which ran 657 tasks across 40 simulated apps including Gmail, Slack, Salesforce and HubSpot, Grok 4.5 scored 51.4%, beating both Fable 5 (48.6%) and Opus 4.8 (48.5%) — and it did so at $0.34 per task versus $1.35 and $1.46 (Yellow, 2026). In finance, the benchmark's hardest category, it led at 71%. It also took the top spot on Harvey's Legal Agent Benchmark (MarkTechPost, 2026).

One caveat worth flagging for enterprise buyers: Grok 4.5 logged 0.63 guardrail violations per task, above Opus 4.8's 0.55 (Yellow, 2026) — a real consideration for regulated, high-stakes workflows.

Why does training on Cursor matter?

Most models learn from static code. Grok 4.5 learned from how developers actually work. After the Cursor tie-up, SpaceXAI trained on trillions of tokens of live editor data — prompts, iterations, tool calls and recoveries from mistakes.

"Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools" — capturing "both existing software as well as developer-agent interactions, capturing how developers work and how agents interact with their environments," Cursor wrote in its launch note (Cursor, 2026).

That focus on the workflow rather than the artefact is the plausible reason a "one-generation-behind" model can still top agent benchmarks: agentic competence isn't the same skill as one-shot code generation. To its credit, SpaceXAI also disclosed a contamination issue — an earlier snapshot of the Cursor codebase was accidentally included in training, inflating its CursorBench result (Cursor, 2026) — a transparency point in an industry that often stays quiet on such things.

What does cheap frontier AI mean for founders and builders?

This is the real story. When frontier-grade agents cost pennies per task, the bottleneck for an AI product stops being model quality and becomes execution economics — cost and speed per completed outcome. Founders who spent 2024–25 rationing tokens can now run agents in loops that were previously uneconomic.

Analysts frame the shift bluntly. "Enterprises are hitting a wall with AI ROI," said Neil Shah of Counterpoint Research. "The massive token consumption required by autonomous agents and coding is causing bill shocks, turning AI adoption into an expensive, one-way street" (InfoWorld, 2026). Grok 4.5 attacks exactly that wall.

The nuance buyers should hold onto: benchmark rank is not the same as value. "Enterprise buyers should focus on cost per successful outcome rather than cost per token," said Forrester's Biswajeet Mahapatra (InfoWorld, 2026). Or, as Omdia's Lian Jye Su put it: "We are living in the era where token consumption is seen as the ultimate value creation but the true value still lies in actual job completion" (InfoWorld, 2026).

The strategic read is familiar: get close enough on performance, then win on price (The Decoder, 2026). For a founder shipping an agent product, "close enough at a quarter of the cost" is often the better trade than "best-in-class and unaffordable at scale."

Who should switch to Grok 4.5 — and who shouldn't?

Strong fit: agentic coding tools, autonomous workflow products, high-volume knowledge-work automation, and any team where per-task cost — not peak capability — is the constraint. If you run models in loops, the token efficiency alone can reshape your margins.

Think twice: teams needing the absolute top of pure coding benchmarks (Fable 5 still leads there), and regulated workflows where the higher guardrail-violation rate is a genuine risk until mitigated. As with any launch, benchmark it on your tasks before you migrate — vendor charts are a starting point, not a verdict.

Visual suggestions

• Pricing bar chart — Alt-text: "Bar chart comparing input and output token prices for Grok 4.5, Opus 4.8, GPT-5.5 and Fable 5." Caption: Grok 4.5 undercuts every current frontier flagship on both input and output pricing.
• AutomationBench leaderboard — Alt-text: "Leaderboard showing Grok 4.5 at 51.4% ahead of Fable 5 and Opus 4.8 on AutomationBench-AA." Caption: On agentic tasks, Grok 4.5 edges the field — at a fraction of the cost per task.
• Cost-per-task comparison — Alt-text: "Chart of cost per coding task: Grok 4.5 $2.49, GPT-5.5 $5.07, Fable 5 $11.80." Caption: Lower rates plus fewer tokens compound into a large real-world cost gap.
• Token-efficiency graphic — Alt-text: "Diagram showing Grok 4.5 using 4.2 times fewer output tokens than Opus 4.8 on SWE-Bench Pro." Caption: Frugal reasoning is the quiet advantage behind the price story.

Frequently asked questions

When was Grok 4.5 released?

SpaceXAI released Grok 4.5 to developers on 8 July 2026 via Grok Build, Cursor and the API, then opened public access on grok.com and the X app on 9 July 2026 (MarkTechPost, 2026).

How much does Grok 4.5 cost?

Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, with cached input around $0.50 per million — roughly a third of Opus 4.8 and a fifth of Fable 5 on output (The Decoder, 2026).

Is Grok 4.5 better than Opus 4.8?

It depends on the task. Grok 4.5 beats Opus 4.8 on the AutomationBench-AA agent benchmark (51.4% vs 48.5%) but trails Fable 5 on pure coding scores like SWE-Bench Pro. Its clearest win is cost per completed task (Yellow, 2026).

Why is Grok 4.5 so much cheaper?

It combines low per-token pricing with high token efficiency, using about 4.2× fewer output tokens than Opus 4.8 on SWE-Bench Pro. Lower rates and lower consumption compound, so per-task costs fall dramatically (MarkTechPost, 2026).

What was Grok 4.5 trained on?

It was trained alongside Cursor on trillions of tokens of real developer-agent interactions — how developers prompt, iterate and use tools — plus STEM tasks and research papers, on tens of thousands of NVIDIA GB300 GPUs (Cursor, 2026).

Should I switch my agent product to Grok 4.5?

If your constraint is cost per task rather than peak benchmark scores, it's a strong candidate — especially for high-volume agentic loops. Test it on your own workloads first, and weigh its higher guardrail-violation rate for regulated use cases (Yellow, 2026).

The bottom line

Grok 4.5 doesn't win by being the smartest model in the room — it wins by being frontier-enough at a price that changes what founders can afford to build. When agents cost pennies per task, profitability shifts from who trained the biggest model to who executes fastest and cheapest per completed outcome. That's a different game, and it favours builders.

Want the next frontier release broken down like this — signal, not hype? Subscribe to the Radar briefing and get launch analysis, real benchmarks and the founder-economics angle in your inbox the day it drops.