Artificial Intelligence

Grok 4.5 Loses the Benchmarks and Wins the Bill

xAI launched Grok 4.5 at 2 and 6 dollars per million tokens, a third of Opus, and it finishes tasks with 4.2 times fewer output tokens. It loses the benchmarks and wins the bill.

Artificial IntelligenceBy Servola Tech Desk2026-07-093 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Grok 4.5 Loses the Benchmarks and Wins the Bill

Key takeaways

On 8 July 2026 xAI released Grok 4.5 at 2 dollars per million input tokens and 6 per million output, roughly a third of Opus 4.8 at 5 and 25, and it resolves a coding task with about 15,954 output tokens against 67,020 for Opus, so the model that loses the top benchmarks can still be the cheaper way to finish the work.
Grok 4.5 trails the leaders on hard software benchmarks, scoring 64.7 percent on SWE-Bench Pro against Fable 5 at 80.4 and 53 percent on DeepSWE 1.1 against 70, so it is not the smartest model on the board and xAI does not claim it is.
The number that decides a budget is cost per completed task, not price per token or leaderboard rank, and on that number a cheaper, more token-efficient model beats a smarter one for high-volume, well-scoped work.
The practical move is a two-model stack: route bulk, well-defined jobs to the cheap efficient model and reserve the premium model for the hard tail, then measure the blended bill rather than picking one model for everything.

What xAI actually shipped

xAI released Grok 4.5 on 8 July 2026, calling it an Opus-class model and making it the default in Grok Build, inside Cursor on every plan, and on the xAI console. It is built on the company V9 foundation model, reported at around 1.5 trillion parameters, and its training was supplemented with real developer session data from Cursor, which is where the coding focus comes from. It is priced at 2 dollars per million input tokens and 6 dollars per million output tokens and serves at roughly 80 tokens a second.

The comparison xAI leads with is not that Grok 4.5 is the smartest model. It is that the model finishes tasks using far fewer tokens. On SWE-Bench Pro xAI reports an average of 15,954 output tokens per resolved task against 67,020 for Opus 4.8, a 4.2 times gap. Put together with a lower per-token price, that is the whole pitch: the same tier of work for a fraction of the bill.

The benchmark you should be reading is cost per task

Grok 4.5 loses the benchmarks that get quoted in headlines. On SWE-Bench Pro it scores 64.7 percent to Fable 5 at 80.4, on DeepSWE 1.1 it manages 53 percent against 70, and on Terminal-Bench 2.1 it lands at 83.3 against a cluster near 84. Read only that column and you would pass on it. But a leaderboard rank is not a budget line.

Put price and token efficiency together and the picture inverts. Grok 4.5 at 2 and 6 dollars per million tokens sits against Opus at 5 and 25, GPT-5.5 at 5 and 30, and Fable 5 at 10 and 50, and then you multiply by the tokens each model burns to finish the job. A model a few points behind on a benchmark but three times cheaper per token and four times leaner per task is not a worse buy for most work. It answers a different question than the benchmark does.

What a European operator should do

The mistake is to pick one model for everything. High-volume, well-scoped jobs, ticket triage, log summarising, first-draft code, routine classification, do not need the top of the board, and paying Fable 5 or Opus rates for them is a leak. For a UK or EU support desk moving millions of tokens a day, the difference between output at about 5.50 euros and about 23 euros per million tokens is the difference between a rounding error and a real monthly line.

Run a two-model stack and measure the blended bill. Route the bulk work to the cheap, token-efficient model, keep a premium model on call for the hard tail where accuracy pays for itself, and log cost per completed task rather than per token. Grok EU footprint is worth a check first: X and Grok are already under a Digital Services Act probe and a GDPR inquiry led from Dublin, so the procurement question is not only price but where your data sits and whether the vendor compliance file is in order.

Frequently asked questions

How much cheaper is Grok 4.5 than Opus?

On list price Grok 4.5 is 2 dollars per million input tokens and 6 per million output, against Opus 4.8 at 5 and 25, so roughly a third of the output price. xAI also reports Grok 4.5 uses about 15,954 output tokens to resolve a SWE-Bench Pro task against 67,020 for Opus, so the gap in cost per finished task is wider than the per-token gap alone.

Is Grok 4.5 the best model available?

No, and xAI does not claim it is. It trails Fable 5 and GPT-5.5 on hard software benchmarks such as SWE-Bench Pro and DeepSWE 1.1. Its argument is cost and token efficiency for a given tier of work, not top rank.

Should a European business switch to Grok 4.5?

For high-volume, well-scoped work it can cut the model bill sharply, so it is worth testing as the bulk-work model in a two-model stack. Before committing, check the EU angle: X and Grok face a Digital Services Act investigation and a GDPR inquiry, so confirm where your data is processed and what contractual terms apply.

The frontier race used to have one scoreboard and it rewarded the smartest model. It now has two, and the second one is priced in tokens. For a business the winning model is not the one at the top of the benchmark, it is the one that clears your actual workload at the lowest total bill, and for a large share of that workload the cheaper, leaner model is now the correct answer.

xAIGrokGrok 4.5AI pricingOpusToken efficiency

DeepSeek Charges Double While Europe Works

DeepSeek V4 launches in mid-July with peak-hour pricing that doubles token rates from 14:00 to 18:00 Beijing time. That window is 08:00 to 12:00 in Frankfurt and 07:00 to 11:00 in London, so European teams pay the surcharge during their own working morning. The deepseek-chat and deepseek-reasoner names retire on 24 July.

2 min read

Artificial Intelligence

GPT-5.6 Terra Matches GPT-5.5 at Half the Price

OpenAI previewed GPT-5.6 Sol, Terra and Luna. Terra matches GPT-5.5 on capability at roughly half the token price, but the new models are locked to preview partners for now.

3 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.