GPT-5.6, Grok 4.5, Muse Spark: Three Efficiency Claims This Week, Zero Common Yardstick

Three model launches landed in the same 48 hours this week, and all three came with an efficiency claim instead of a benchmark table. OpenAI said GPT-5.6 Sol is 54% more token efficient on agentic coding. xAI said Grok 4.5 gets 4.2x the output-tokens-per-solved-task of Claude Opus 4.8. Meta priced Muse Spark 1.1 at roughly 25% of Anthropic and OpenAI's rates and called it a preview of where the market is going. Read those three numbers next to each other and you'd think you have a clean ranking. You don't, because none of them were measured against the same baseline, on the same task set, or by anyone outside the company making the claim.

The three claims, as stated

• OpenAI / GPT-5.6 Sol: Altman told CNBC the model is 54% more token efficient on agentic coding than its predecessor. OpenAI's own comparison point was Claude Mythos Preview on ExploitBench, where Sol matched performance while generating about one-third the output tokens. GPT-5.6 shipped in three sizes: Sol (frontier), Terra (balanced), Luna (fast), priced from $1/$6 per million tokens on Luna.
• xAI / Grok 4.5: priced at $2 input / $6 output per million tokens, with xAI claiming a 4.2x advantage over Opus 4.8 in output tokens per solved coding task. That comparison is against Opus 4.8's list price of $5/$25 per million tokens, which is xAI's own chosen reference point.
• Meta / Muse Spark 1.1: API priced at $1.25 input / $4.25 output per million tokens, which Zuckerberg framed as about a quarter of Anthropic and OpenAI's rates. The fine print: Muse Spark is a reasoning model, and its chain-of-thought tokens bill at the output rate, not a discounted one. That detail doesn't show up in the headline price.

Why the numbers don't stack

Every one of these figures is a company measuring itself against a benchmark and a rival model that it picked. OpenAI chose ExploitBench and Claude Mythos Preview. xAI chose Opus 4.8's list price as the denominator. Meta's 25% figure is a sticker-price ratio that ignores how reasoning tokens actually get billed. None of the three used the same task, the same harness, or an outside grader. Stack the claims and you get the illusion of a ranking, when what you actually have is three companies answering three different questions and printing the results in the same units.

Meta Slashes AI Prices by 75%, Undercutting OpenAI and Anthropic with Muse Spark 1.1 - BigGo Finance headline, July 10, 2026

That's a real headline from this week, and it's not wrong as a sticker-price comparison. But a 75% discount on list price and a 54% efficiency gain on a self-chosen benchmark are not the same kind of number, and neither tells you what a given task will actually cost once a model runs long or short on a specific prompt. We've made this point before about sticker price versus delivered cost, and this week is the clearest version of it yet: three vendors, three different self-reported efficiency metrics, released within two days of each other.

What would actually make the claims comparable

The fix isn't complicated, it's just something none of the three companies has an incentive to do: run the same one-shot prompt through all three models, on the same day, with full output-token counts visible, and let people who don't know which output came from which model decide which result they'd actually want to ship. That's the whole premise of running an arena instead of reading a launch post. It's also why effort level matters more than model identity, something we've written about at length: cranking a model's thinking budget up or down changes its output-token count dramatically without changing which model it is, so a vendor's efficient claim needs to specify which effort setting produced it, or the number is close to meaningless.

None of this means the three launches are bad models. GPT-5.6 Sol, Grok 4.5, and Muse Spark 1.1 are all real, shipped, and priced aggressively enough that the next few months of coding-assistant pricing will move because of them. But aggressive pricing and efficient are marketing words attached to numbers that were chosen by the party with the most reason to pick a flattering comparison. That's not a criticism specific to any one of the three; it's the default state of a self-reported benchmark, whoever publishes it.

Where the arena fits

This is exactly the gap a blind, live arena is built to close. Run the same prompt through GPT-5.6 Sol, Grok 4.5, and Muse Spark 1.1 side by side, show the raw output-token count for each response, and let votes land without anyone knowing which model produced which answer. No one gets to pick their own benchmark or their own comparison model. If you want to see how these three actually behave on a real task rather than a press release, open the coding arena and run one yourself; the token counts are right there next to the outputs. Readers following the same story in Polish can track it at nowosci.ai, which covers the same week of launches from the other side of the Atlantic.

The honest version of this week's news isn't Meta is 75% cheaper or GPT-5.6 is 54% more efficient. It's that three well-funded labs all decided, independently, that the fight to win right now is on price and token count rather than raw capability, and each picked the comparison that made their own number look best. That's worth watching. It's not worth quoting as a fact until someone runs it blind.