For anyone shipping a real product on top of a US frontier API, the cheapest line item on a cloud bill just changed hands. SpaceXAI released Grok 4.5 to the public on July 9, 2026, with a list price of $2 per million input tokens and $6 per million output tokens - the first release from the company since it went public several weeks ago. That single line is why the rest of this article exists, because the rate it sets resets the floor every other frontier lab is now pricing against.
What $2 and $6 actually buy
The list price is the part that matters. Per the TechCrunch write-up, SpaceXAI is selling Grok 4.5 at “$2 per million input tokens and $6 per million output tokens.” The same rate is reflected on the OpenRouter Grok 4.5 listing, which describes the model as “SpaceXAI’s smartest model with frontier performance on coding, knowledge work, and STEM” and confirms the $2 input / $6 output rate as the published price across providers.
Elon Musk’s own framing, quoted by TechCrunch from his X post, was direct: “It is an Opus-class model, but faster, more token-efficient and lower cost.” A second post the same day said, “Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster. The combination of capability, faster speed and lower cost is what makes it competitive.” That is a public comparison from the person selling it, not a third-party benchmark - and it is the comparison every other frontier lab now has to either accept or rebut.
The second claim worth holding on to is the token-efficiency number. TechCrunch reports that SpaceXAI is positioning Grok 4.5 as having “‘twice greater token efficiency’ than other leading models,” and ties it directly to the unit-economics argument: “If it carries through to real-world use cases, that efficiency would be a big advantage for SpaceXAI, since the cost of tokens has been a growing concern for AI consumers.” Token efficiency is a real lever, because output tokens are where most production bills land and a model that uses half as many to do the same work costs half as much before the list rate even matters.
How it stacks up against the other US frontier labs
The headline rate only matters in context, and it lands on top of the API pricing collapse we traced back in March and the cost crackdown we documented earlier this month. The cleanest comparison set, drawn from the same TechCrunch article and corroborated against the providers’ own public pricing, looks like this:
- Grok 4.5: $2 input / $6 output per million tokens.
- Claude Opus 4.7: $5 input / $25 output per million tokens, per TechCrunch and confirmed by the OpenRouter Claude Opus 4.7 listing, which carries a 1M-token context window and a release date of April 16, 2026.
- OpenAI’s “Sol” tier (which OpenAI’s pricing page lists as GPT-5.5): $5 input / $30 output per million tokens, per TechCrunch and confirmed by OpenAI’s own model documentation.
On input, Grok 4.5 is 2.5x cheaper than Opus 4.7 and 2.5x cheaper than GPT-5.5. On output, it is roughly 4.2x cheaper than Opus 4.7 and 5x cheaper than GPT-5.5. The output gap is the one that matters for production workloads, because most real applications spend more on the model’s response than on the prompt.
There is one honest counterpoint the article itself surfaces. OpenAI’s cheapest frontier tier - “Luna” in TechCrunch’s reporting - is listed at $1 per million input and $6 per million output. That is actually cheaper than Grok 4.5 on input, and identical on output. The framing of “cheapest US frontier API” depends entirely on which OpenAI tier you compare to: against “Sol”/GPT-5.5, Grok 4.5 wins on both axes; against the cheapest OpenAI tier, it wins on output and loses on input by a dollar.
The other caveat is that the comparison is a list-price comparison, not a like-for-like capability comparison. TechCrunch’s own summary of the released benchmark numbers is that Grok 4.5 is “competitive with other top models from SpaceXAI competitors, although just short of best-in-class.” The “Opus-class” framing is Musk’s, not an independent eval.
What This Means
For an intelligibberish reader building a product, the practical implications are three.
First, the unit-economics case for the new rate is real even before the “twice greater token efficiency” claim is factored in. For a workload that splits input and output tokens roughly 50/50 - which is closer to how most chatty agentic workloads actually bill - the simple arithmetic of the cited rates works out to a Grok 4.5 line item of $8 per million tokens ($2 plus $6) against $30 per million on Opus 4.7 and $35 per million on GPT-5.5. That is a several-times-per-million-token gap, applied to whichever line of the cloud bill output tokens sit on, and it does not depend on whether the “twice greater token efficiency” claim holds up.
Second, “Opus-class” is the only part of the marketing that matters to watch. Token-efficiency claims tend to hold on simple chat but compress or invert on long-context workloads, where a model has to keep re-reading a large prompt or carry a long conversation history. The same TechCrunch piece acknowledges the open question with the line “if Grok’s capabilities match SpaceXAI’s rhetoric.” Until an independent benchmark against Opus 4.7 and GPT-5.5 lands - one that uses real long-context tasks and not the press-quoted headline number - the cheapest-Opus-class claim is a sales position, not a measured fact.
Third, this is the floor that has to be answered, not the ceiling that has to be matched. OpenAI’s planned GPT 5.6 release the following day, GPT-5.4 mini sitting at $0.75 / $4.50 on OpenAI’s own pricing page, and Anthropic’s enterprise insulation from the consumer side all mean the next move is a public rate-card change, not a quiet press release. The pricing floor for US frontier APIs moved on July 8, and the labs that have to either match it or explain why they are not matching it now have a deadline measured in news cycles, not quarters.
The Bottom Line
Grok 4.5 undercuts Claude Opus 4.7 and OpenAI’s top GPT-5.5 tier on both input and output per million tokens, and is being sold on a public claim of “twice greater token efficiency” by a company whose owner is calling it “Opus-class.” The list price is real, verified, and below both benchmarks. The capability claim is the founder’s, not an independent benchmark’s, and it is the part to test before betting a production workload on it. OpenAI’s cheapest tier remains lower on input alone.