SpaceXAI shipped Grok 4.5 to its public API and consumer endpoints on Tuesday, July 8, 2026. Elon Musk described the model as "Opus-class" in a post on X the same morning. The launch ended a private beta that began in mid-May and included, by SpaceXAI's own count, fewer than 40,000 external testers. The model is now generally available. The benchmark stack the company published the same day is the first public third-party check on that claim.
The Launch: Exiting the Beta Cocoon
Grok 4.5's private beta opened on May 14, 2026, according to SpaceXAI's developer changelog. The beta was invite-only, distributed through the X Premium+ tier and a small set of API keys issued to researchers and a handful of named startups. The number 40,000 is the figure SpaceXAI used in its launch post. It is also the figure the company has not, as of Tuesday evening, broken down by use case, geography, or evaluation methodology.
What is generally available as of Tuesday: a 256k context window, function calling, structured outputs, image and document inputs, and native tool use. What is not generally available: video input, real-time voice, and a fine-tuning API. The fine-tuning API is the omission that matters. Anthropic offers it for Sonnet and Haiku. OpenAI offers it for GPT-4o and the o-series reasoning models. SpaceXAI's launch post does not commit to a date. The omission is the kind of thing that does not matter to a consumer and matters considerably to an enterprise procurement officer.
The release notes describe Grok 4.5 as a "frontier-tier reasoning and coding model." The phrase "frontier-tier" is doing the same work the phrase "evaluation" did in last week's defense reporting. It implies a position on a curve. It does not name the curve. The next section names the curve.
"Opus-Class" — Reading the Receipt and the Stub
"Opus-class" is not a benchmark. It is a positioning. Anthropic's Opus product line sits at the top of the company's three-tier structure: Haiku, Sonnet, Opus. Sonnet is the workhorse. Opus is the one Anthropic prices at the top of the public market and the one it deploys for its most difficult evaluations. Calling a model "Opus-class" is a claim that it competes in the same tier, on the same kinds of tasks, at the same price band. It is a useful claim because it gives customers a frame of reference. It is a dangerous claim because it concedes the comparison.
Musk's post is the receipt. The receipt-stub is what the post does not contain: a head-to-head table. SpaceXAI's launch materials include a benchmark sheet. The sheet does not include a row for Claude Opus 4.1. It includes rows for GPT-4.5, Gemini 2.5 Pro, and a set of internal "Grok 3.5" baselines. The omission is consistent with a launch posture that is comparative without being confrontational. The omission is also why the model is being called Opus-class rather than called better than Opus.
Anthropic has not, as of Tuesday afternoon, issued a public response. The company's standard practice is to not respond. The practice is not new. It was the same response, or non-response, when OpenAI shipped GPT-4.5 and described that model as a frontier system. The non-response is itself a position. It says: we do not need to dignify the comparison with a chart.
What the Public Benchmark Sheet Actually Shows
SpaceXAI's published numbers, summarized: 88.4% on MMLU-Pro, 74.1% on GPQA Diamond, 64.9% on SWE-bench Verified, 81.2% on HumanEval-Plus, and 91.3% on a private internal coding benchmark the company calls "GrokCode." The first three are public. HumanEval-Plus is public. GrokCode is not. The 91.3% number is therefore the one figure on the sheet that cannot be independently checked, and it is the figure SpaceXAI leads its announcement with.
The 64.9% on SWE-bench Verified is the number worth dwelling on. It places Grok 4.5 between Claude Sonnet 4 (61.2%) and Claude Opus 4.1 (69.4%) on real-world coding tasks. The number does not place Grok 4.5 in Opus territory. It places Grok 4.5 in Sonnet territory with an asterisk. The asterisk is the model's tool use, which SpaceXAI enabled for the SWE-bench run and which Anthropic does not enable by default for its public evaluations. Tool use, in 2026, is the difference between a 64.9 and a 69.4.
An independent evaluation, posted to arXiv on the same day and authored by a group at Stanford and MIT, runs Grok 4.5 on a held-out subset of the same benchmarks and reports numbers that are 1.8 to 3.4 percentage points lower than SpaceXAI's. The discrepancy is consistent with the standard gap between internal and external evaluation. The paper, which is available on arXiv, also notes that the model exhibits the same jailbreak susceptibility as its predecessor, with a 6.1% higher failure rate on the HarmBench prompt set. The figure was not in the launch materials.
SpaceXAI's Stack: X, Tesla, and the Compute Tailwind
The training corpus is the part of the launch that SpaceXAI describes with the least specificity. The company confirms continued pre-training on the X firehose, including posts, replies, and the full thread graph. It confirms continued use of Tesla's fleet video, drawn from the 6.2 million vehicles shipping telemetry. It does not publish a token count, a data mix, or a deduplication rate. The training cluster is described as "the expanded Colossus-2 footprint in Memphis," with a peak compute figure the launch post declines to specify beyond "north of 200,000 H100-equivalents."
The vertical integration argument is real. The argument is that SpaceXAI can train on data its competitors cannot obtain, on compute its competitors cannot build, at a cadence its competitors cannot match. The argument has been the same since the company's founding. What changed between Grok 4 and Grok 4.5 is the compute: the 200,000-H100 figure is roughly 2.4x the figure Musk cited for the previous generation. The training data did not change proportionally. The compute did. The implicit claim is that scale is doing the work that data was previously doing.
This is the part of the story that makes the triangle possible. It is also the part that does not appear in the marketing copy.
From Duopoly to Triangle
For the eighteen months ending in May 2026, the frontier model market had the structure of a duopoly. OpenAI and Anthropic shipped models that were, by most public measurements, ahead of the rest of the field by a generation. Google shipped Gemini, but the gap between Gemini 2.0 and the OpenAI/Anthropic frontier was measurable and visible. Meta shipped Llama, but the open-weight releases were a tier below. SpaceXAI shipped Grok 3, but Grok 3 was, by the standards of the rest of the field, a step behind.
Grok 4.5 is the first SpaceXAI release that closes most of that gap on the public benchmarks. It does not close it on every benchmark. It does not close it on price. It does not close it on the qualitative evaluations that enterprise customers use. It does close the gap enough that the frontier market is, for the first time, structurally a three-corner race. The corners are not symmetric. Anthropic leads on coding and enterprise adoption. OpenAI leads on consumer distribution and the developer ecosystem. SpaceXAI leads on training data vertical integration and on raw compute throughput. The corners are, however, close enough to be called a triangle. A Tuesday-morning digest from New Horizon uses the same word.
Pricing, Latency, and the 5.0 Whisper
API pricing for Grok 4.5 is set at $3.00 per million input tokens and $15.00 per million output tokens, with a batch discount of 25% and a 50% cache discount on repeated prefixes. The price is, at the input end, slightly above Sonnet 4 ($3.00) and below Opus 4.1 ($5.00). At the output end, it is below both. The pricing structure is the most direct reading of the "Opus-class" claim: cheaper than Opus, more expensive than Sonnet, in the band that says "we are not conceding the tier."
Median latency on a 64k context prompt is reported at 410 milliseconds to first token, with an inter-token throughput of 87 tokens per second. Both numbers are competitive. Neither is best-in-class. Anthropic reports 380 milliseconds for Opus 4.1 at the same context length. The 30-millisecond gap will not move enterprise procurement decisions. The price gap might.
The 5.0 whisper is the part of Tuesday's coverage that the launch post does not include. Two SpaceXAI engineers, in posts that were deleted within four hours, referenced "the 5.0 roadmap" and a target internal milestone of Q4 2026. The posts were deleted. The screenshots are in the same New Horizon digest. A 5.0 launch in Q4 2026 would put SpaceXAI roughly six months behind OpenAI's GPT-5 cadence, which has been rumored for a similar window. The triangle, if it holds, will not hold for long at three corners. Tuesday also brought OpenAI's release of a new set of voice models aimed at live conversation, a category SpaceXAI does not yet ship to API. The shape of the triangle changes faster than the labels on its corners. The map, drawn now, is already out of date.