xAI dropped Grok 4.5 yesterday and the benchmark chart is the least interesting part of the announcement. Read the second chart instead. Grok 4.5 charges $2 per million input tokens. Opus 4.8 charges $5. Fable 5 charges $10. Then xAI casually mentioned their model uses 4.2 times fewer tokens than Opus 4.8 on SWE Bench Pro tasks. Do that multiplication and the "premium AI" framing collapses.
Let me say the quiet part out loud: the per-task cost gap on coding is no longer a rounding error. It is the entire story.
The benchmarks are real, the gap is real, and it does not matter
Grok 4.5 is not winning on raw coding accuracy. On SWE Bench Pro it scores 64.7%, behind Fable 5's 80.4% and Opus 4.8's 69.2%. On DeepSWE 1.1, which measures real GitHub issue resolution, it lands at 53% versus Fable 5's 70% and GPT-5.5's 67%. The frontier labs are still ahead on the hard stuff.
But here is the part xAI wants you to internalize: on Terminal Bench 2.1, Grok 4.5 scores 83.3%, almost exactly tied with GPT-5.5 (83.4%) and within one point of Fable 5 (84.3%). On command-line tasks the gap vanishes. And on a real workload, you pay per token, not per benchmark position.
Pricing on coding tasks: the math nobody wants to do
The pricing table from xAI is short and brutal:
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Grok 4.5 | $2 | $6 |
| Opus 4.8 | $5 | $25 |
| GPT-5.5 / 5.6 | $5 | $30 |
| Fable 5 | $10 | $50 |
Now stack token efficiency on top. xAI says Grok 4.5 uses 4.2x fewer output tokens than Opus 4.8 on SWE Bench Pro and serves at 80 tokens per second. A representative SWE Bench Pro task that costs Opus 4.8 around $40 of output tokens costs Grok 4.5 about $3.60, even before you account for the lower input price. The headline "$2 per million" is the cover charge. The real number is what happens when you multiply it by the efficiency multiplier.
The benchmark table in xAI's launch chart does include GLM-5.2 at 62.1% SWE Bench Pro, which is unusual. Open-weights models do not normally appear in frontier lab comparison tables. Worth flagging.
What this actually changes for working developers
I have been running Opus 4.8 in Claude Code at max effort for a while, and the cost ceiling was always the thing I was managing, not the quality ceiling. Grok 4.5 is now available in Cursor and the xAI console, with mid-July EU availability. Cursor is the interesting channel because SpaceX closed the $60B acquisition in mid-June, so the IDE and the model now sit on the same balance sheet.
A few practical implications:
- Background jobs (refactors, test sweeps, repo-wide migrations) used to be cost-prohibitive on Opus. They are now cheap enough on Grok 4.5 that you can actually run them overnight.
- Token efficiency at 80 TPS changes the latency story for streamed code completion, not just batch work.
- The closed-weights tier is converging with the open-weights tier on price-per-task even when it is not converging on raw accuracy. That is the part that should worry the labs charging 5x more.
What it does not change
Grok 4.5 still trails on the hardest coding benchmarks. If you are doing novel architecture work where each token of context is a high-stakes decision, Fable 5 at $50/M output is still going to be the tool of choice for a while. The pricing war is not a quality war. It is a "good enough at 1/10th the cost" war, and that is a different war entirely.
The Chinese labs already figured this out. Zhipu and DeepSeek have been running the same playbook for a year. xAI just imported it into the US frontier tier with a closed model and a CUDA budget measured in tens of thousands of GB300s.
My read
The interesting move is not Grok 4.5 itself. It is what happens to Opus 4.8 pricing over the next 90 days. If Anthropic does not reprice, every agentic workflow that is not on the absolute hardest benchmark is going to quietly migrate to cheaper models. And if Anthropic does reprice, Fable 5 has to be next, which puts pressure on the entire closed-weights tier.
Either way, "$2 per million" is the new floor. Everyone else is now arguing about whether to match it, beat it, or pretend it does not exist.
Source: the-decoder.com