SpaceXAI, Elon Musk's AI company, released Grok 4.5 on July 8, 2026. The pitch is unusually blunt: it is not the best model on the leaderboard, and that may not matter. As The Decoder put it, Grok 4.5 trails the top coding models on raw benchmark scores while costing a fraction of what they charge and burning far fewer tokens to reach an answer. In a market where the frontier labs keep chasing the highest number, SpaceXAI is betting that price and efficiency win more real work than a few points of benchmark headroom.
- Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, well under Opus 4.8 ($5/$25) and Fable 5 ($10/$50).
- On SWE Bench Pro it uses 4.2 times fewer tokens than Opus 4.8, which compounds the price advantage on long agentic runs.
- It lags the leaders on coding benchmarks like DeepSWE and SWE Bench Pro, so the case rests on cost-per-task rather than peak capability.
What SpaceXAI shipped
Grok 4.5 arrived on a Wednesday with public access opening the following day, according to TechCrunch, which noted that SpaceXAI itself went public only several weeks earlier. The company is positioning the model as a general-purpose workhorse rather than a specialist, aiming it at coding and app-building alongside office work, research and everyday knowledge tasks. The through-line in the marketing is value: a model that gets close enough to the best while costing much less to run.
That framing is a deliberate contrast with how OpenAI and Anthropic have been selling their flagships. Where those labs lead with capability and let price follow, SpaceXAI is leading with price and asking buyers to decide whether the capability gap is large enough to justify paying two or three times more. It is a classic challenger move, and the timing sharpened it: the release landed a day before OpenAI's planned GPT-5.6 launch, guaranteeing Grok 4.5 a share of the same news cycle.
The pricing gap that defines Grok 4.5
The numbers are the story. Grok 4.5 (July 2026) charges $2 per million input tokens and $6 per million output tokens. Anthropic's Opus 4.8 sits at $5 in and $25 out, and Fable 5 runs $10 in and $50 out. On output pricing, the dimension that dominates cost for anything that generates long code or documents, Grok 4.5 is roughly four times cheaper than Opus 4.8 and more than eight times cheaper than Fable 5.
TechCrunch framed the same gap against Anthropic's Opus 4.7 at $5/$25 and OpenAI's Sol model at $5/$30, which puts Grok 4.5 at well under half the input price and a quarter or less of the output price of its nearest rivals. For a team running an AI coding agent that chews through millions of tokens a day, that difference is not a rounding error. It is the line between a tool that is affordable to run at scale and one that gets rationed.
A rough worked example makes the gap concrete. Say an automated task reads 500,000 tokens of context and writes 200,000 tokens of code across a multi-step run. On Fable 5's $10/$50 rates that costs about $5 in input plus $10 in output, roughly $15. On Opus 4.8's $5/$25 it lands near $2.50 plus $5, about $7.50. On Grok 4.5's $2/$6 the same token counts cost $1 plus $1.20, around $2.20. That is before the separate token-efficiency claim enters the picture, which would push Grok's real number lower still. Run that task ten thousand times a month and the per-model bills diverge into very different budget categories.
The output rate is the number to watch most closely. Generation dominates cost for anything that produces code, drafts or long reasoning traces, because output tokens are priced several times higher than input tokens across every vendor. Grok 4.5's $6 output rate against Fable 5's $50 is the single widest gap in the comparison, and it is exactly the dimension that hurts most when a model is asked to write a lot rather than just read a lot. For code-generation workloads specifically, that is close to a best-case shape for SpaceXAI's pricing argument.
Musk's own framing, quoted by TechCrunch, leans entirely on that value angle. He added that SpaceXAI's internal assessment puts Grok 4.5 as roughly comparable to Opus 4.7 but much faster. Note the careful comparison: Opus 4.7, not the newer Opus 4.8. The company is claiming parity with the previous generation of Anthropic's flagship, not the current one, which is an honest signal about where the model actually lands.
Token efficiency and the 4.2x claim
Price per token is only half of what a task costs. The other half is how many tokens the model spends getting there, and this is where Grok 4.5 makes its most interesting claim. On SWE Bench Pro, a software-engineering benchmark, The Decoder reported that Grok 4.5 uses 4.2 times fewer tokens than Opus 4.8 to complete the same work. SpaceXAI describes this more broadly as roughly twice the token efficiency of leading models.
Token efficiency multiplies with the per-token price to produce the real bill. If a model is cheaper per token and also needs fewer tokens, the two effects stack. A task that costs a dollar on Opus 4.8 could plausibly cost a small fraction of that on Grok 4.5 once both the lower rate and the reduced token count are accounted for. For agentic workloads that loop through many steps, each generating and consuming tokens, that compounding is the whole argument.
The model runs at about 80 tokens per second, which puts it in a comfortable range for interactive coding and chat without being the fastest option available. Speed matters for developer experience, but SpaceXAI's real emphasis is on the economics of long runs rather than the latency of a single reply. The company also disclosed that its reinforcement-learning stage covered hundreds of thousands of tasks weighted heavily toward software engineering, which explains why the model's efficiency story is strongest on coding benchmarks specifically.
Why would one model need fewer tokens than another for the same task? Part of it is training that discourages rambling reasoning traces, so the model reaches a conclusion with less internal chatter. Part of it is how the model is tuned to plan before it acts rather than exploring many dead ends. A model that solves a coding problem in three well-chosen steps spends far fewer tokens than one that tries eight approaches before landing on the right one, even if both eventually succeed. The 4.2x figure on SWE Bench Pro implies Grok 4.5 does markedly less of that exploratory wandering than Opus 4.8 on the same problems, which is as much a property of the training recipe as of the architecture.
The catch is that token efficiency is task-dependent, and a single benchmark cannot promise it holds everywhere. SWE Bench Pro is a curated set of software problems; a model tuned to be terse on those may or may not be equally frugal on open-ended research, long-form writing or multi-turn conversation. SpaceXAI's broader claim of roughly twice the efficiency covers more ground but is correspondingly softer. The honest read is that the coding number is the one with a benchmark behind it, and buyers whose workloads look like SWE Bench Pro have the strongest reason to expect the savings to transfer.
Where Grok 4.5 sits on the benchmarks
The benchmark table is where the honesty of the pitch shows. On DeepSWE 1.1, Grok 4.5 scored 53% against Fable 5's 70% and GPT-5.5's 67%, a clear gap. On SWE Bench Pro it reached 64.7%, ahead of GPT-5.5 at 58.6% but behind Fable 5 at 80.4%. On Terminal Bench 2.1 the field bunched tightly: Grok 4.5 at 83.3%, Fable 5 at 84.3%, and GPT-5.5 at 83.4%, effectively a three-way tie.
So the picture is mixed rather than dominant. On some tasks Grok 4.5 is genuinely competitive. On the hardest coding benchmark it trails the leader by a wide margin, and it takes the top spot on none of them. TechCrunch summarised it as landing just short of best-in-class, which is a fair reading. The company is not claiming the crown. It is claiming that for most practical work, the difference between 64.7% and 80.4% matters less than paying a quarter of the price.
It helps to know what these benchmarks actually stress. SWE Bench Pro asks a model to resolve real software issues in real repositories, which rewards long-horizon reasoning and the ability to hold a large codebase in context. Terminal Bench measures how well a model operates a command line to accomplish tasks, closer to raw tool use than to abstract reasoning, which is why the field bunches so tightly there. DeepSWE sits in between. The pattern in Grok 4.5's scores suggests a model that is solid at operating tools and following procedures but gives up ground on the deepest, most open-ended engineering problems, exactly the tasks where the premium models still earn their premium.
All efficiency and benchmark figures here come from SpaceXAI's own release materials as reported by The Decoder and TechCrunch. Vendor benchmarks favour the vendor, and the token-efficiency and cost comparisons in particular have not been independently reproduced. The pricing is verifiable; the performance claims await outside testing.
Whether that trade lands depends entirely on the buyer. For a startup running an autonomous coding agent on a tight budget, a model that is 80% as good at a quarter of the cost is an obvious win. For a team where a single missed edge case is expensive, the extra points at the top of the benchmark may be worth every dollar. Grok 4.5 does not try to be for everyone; it tries to be the rational default for the large middle of the market that cares more about cost per task than about the last few points of capability.
Trained on tens of thousands of GB300 GPUs
The efficiency did not come from nowhere. The Decoder reported that Grok 4.5 was trained on tens of thousands of Nvidia GB300 GPUs, the current top of Nvidia's data-center line, with heavy data filtering, deduplication and domain-specific selection during pretraining. That emphasis on data curation rather than raw scale is consistent with a model optimised for doing more with fewer tokens rather than simply being larger.
The reinforcement-learning phase leaned hard into software engineering, covering hundreds of thousands of tasks. That focus is visible in the results: the model's strongest showings and its efficiency advantage both cluster around coding, while its general-knowledge positioning is broader and less benchmarked. SpaceXAI is being upfront that this is, at heart, a model tuned for building things, with the wider capabilities as supporting range rather than the headline.
Access at launch runs through Grok Build, the Cursor coding environment and the xAI console, with Microsoft Office plugins for Word, PowerPoint, and Excel already live. European availability is targeted for mid-July, a staggered rollout that has become standard as labs navigate the EU's regulatory posture on frontier models. The Office integration is a notable tell: it signals that SpaceXAI wants Grok 4.5 in front of ordinary knowledge workers, not just developers running it through an API.
Landing the release a day ahead of GPT-5.6 was almost certainly deliberate. A cheaper challenger benefits from being measured against a rival while that rival is fresh in buyers' minds, and forcing the comparison on its own terms lets SpaceXAI steer the conversation toward price rather than peak capability. It is the kind of timing that only makes sense if a company is confident its story is about value, because a model chasing the top benchmark score would rather not be reviewed side by side with the newest flagship from a better-funded lab.
The Opus-class framing and what it signals
Musk's "Opus-class" description is doing careful work. By benchmarking against Opus 4.7 rather than the newer Opus 4.8, SpaceXAI sets an expectation the model can actually meet, then adds the qualifiers that matter for its pitch: faster, more token-efficient, lower cost. It is a more disciplined message than the usual claim of unqualified supremacy, and it fits a company that knows its benchmark scores do not lead the field.
There is a broader pattern here worth naming. The frontier labs have spent 2026 pushing capability up and price up with it, with Fable 5 at $10/$50 sitting at the expensive end. SpaceXAI is testing whether a large segment of buyers will trade a measurable capability gap for a large cost saving. If they do, it pressures the premium labs to justify their pricing or cut it. The same dynamic is playing out with cheap open-weight models from Chinese labs steadily gaining share on cost, so Grok 4.5 is one move in a wider repricing of the market. Our running coverage of the model landscape tracks how fast that shift is moving.
The competitive risk for SpaceXAI is that value is the easiest position to attack. Anthropic and OpenAI can cut prices on a spreadsheet far faster than SpaceXAI can close a benchmark gap through training, and both have deeper pockets to absorb thinner margins. A model whose main advantage is being cheaper lives or dies on staying cheaper, and the incumbents have every incentive to erase that edge the moment it starts costing them customers. What makes Grok 4.5's position more durable than a simple discount is the token-efficiency claim, because using fewer tokens is a structural cost advantage a rival cannot match just by lowering its rate card. If that efficiency proves real and holds across workloads, it is the part of the pitch the premium labs cannot answer with a price cut alone, and the part worth watching as independent testers get their hands on the model.
How the value bet plays out
Grok 4.5 is a bet that the AI market is maturing past the phase where buyers pay any price for the top model. On benchmarks it is a solid second tier, clearly behind Fable 5 on the hardest coding tasks and competitive on the easier ones. On economics it is aggressive, undercutting its rivals on both the per-token rate and the token count a task requires, which is the combination that actually determines the bill for anyone running agents at scale.
The claims that matter most, the 4.2x token-efficiency figure and the cost-per-task advantage, are vendor-reported and still need independent confirmation. The pricing, at least, is a hard fact anyone can check. If the efficiency holds up under outside testing, Grok 4.5 becomes one of the strongest value propositions in the market, and the premium labs will have to answer it on price rather than just on capability. If it does not, the model is a competent, cheap option that trails the leaders and knows it. Either way, SpaceXAI has reframed the question from which model is best to which model is worth the money, and that is a harder question for the expensive incumbents to wave away.