Grok 4.5 Launch: Pricing, Benchmarks, Opus Claim

SpaceXAI has launched Grok 4.5, its newest flagship model, positioning it explicitly as an “Opus-class” system that matches Anthropic’s Claude Opus 4.8 on capability while undercutting it sharply on price. The model went live for developers on July 8, 2026 through Grok Build, inside the Cursor coding editor, and via the SpaceXAI API, with public access opening on grok.com and the X app the following day. It lands on the same crowded week that saw OpenAI’s GPT-5.6 family enter limited preview — one of the tightest clusters of frontier releases the industry has seen.

What Grok 4.5 is

Grok 4.5 is a mixture-of-experts model built on xAI’s 1.5-trillion-parameter V9 foundation and co-trained with Cursor on trillions of tokens of coding data. That training partnership is the through-line of the release: SpaceXAI agreed to acquire Cursor for a reported $60 billion in June 2026, and Grok 4.5 is the first model to ship out of that arrangement, tuned from the start for agentic software work rather than retrofitted for it.

The mixture-of-experts architecture routes each token through only a fraction of the model’s total parameters, which is how xAI keeps inference fast and cheap on a system this large. The company reports throughput of roughly 80 tokens per second, and Grok 4.5 is now the default model inside Grok Build.

On specs, the headline numbers are a 500,000-token context window — notably down from the 1-million-token window of the earlier Grok 4.3 — and pricing of $2 per million input tokens and $6 per million output tokens, with cached input at $0.50 per million, a 75% discount for repeated context. A surcharge applies above 200,000 tokens.

The benchmark story is mixed

Elon Musk framed Grok 4.5 as “Opus-class, faster, more token-efficient, and lower cost.” The benchmarks SpaceXAI chose to publish tell a more nuanced story: Grok 4.5 wins on two of four and loses on two.

• DeepSWE 1.0: Grok 4.5 62.0% vs Opus 4.8 (max) 55.75% — a clear Grok win.
• Terminal-Bench 2.1: Grok 4.5 leads Opus 4.8 on the command-line agentic benchmark.
• DeepSWE 1.1: Grok 4.5 53% vs Opus 4.8 59% — Opus ahead by 6 points.
• SWE-Bench Pro: Grok 4.5 64.7% vs Opus 4.8 69.2% — Opus ahead by 4.5 points.

The pattern is worth reading closely. Grok 4.5 leads on the older DeepSWE 1.0 and on Terminal-Bench, and trails on the newer DeepSWE 1.1 and on SWE-Bench Pro, the software-engineering benchmark where Anthropic’s models have held an edge for a generation. In other words, Grok is near the frontier without clearing it — a strong second, not a new leader. That is a familiar shape: labs lead with the tests they win, and independent evaluations tend to compress the gaps in both directions.

Where Grok 4.5 actually pulls ahead: token efficiency

The most striking number is not an accuracy score. On SWE-Bench Pro, SpaceXAI reports that Grok 4.5 resolves tasks using an average of 15,954 output tokens, against 67,020 for Opus 4.8 on the same benchmark — a roughly 4.2x gap.

That efficiency is the crux of the value proposition. Even where Grok 4.5 scores a few points lower than Opus on raw success rate, it can reach its answer using a quarter of the output tokens, and output tokens are where the money is. Combined with a $6-per-million output price, the effective cost of completing a coding task on Grok 4.5 can land well below a nominally stronger competitor. For autonomous coding agents that fan a model across thousands of tool calls per task, token economy compounds into a real operating-cost difference — the kind of gap that prompt caching and careful model selection exist to exploit.

The SpaceXAI backdrop

The release also formalizes a corporate rebrand that has been underway since early 2026. SpaceX acquired xAI in February 2026, and the AI products have been migrating under the SpaceX umbrella since — the account and console now carry the SpaceXAI name, and Grok 4.5’s documentation and pricing pages live on that brand. The Colossus 2 supercluster in Memphis, xAI’s training backbone, remains the compute engine behind the V9 foundation.

The distribution strategy is unchanged from the Grok 4.3 playbook: put the model everywhere enterprises and developers already work. Grok 4.5 ships simultaneously in Grok Build, in Cursor across all plan tiers, and through the standard API. Meeting developers inside the tools they already use — rather than asking them to adopt a new surface — is how xAI has consistently tried to convert capability into usage.

How it fits the competitive picture

Grok 4.5 arrives into a market where the price of a given capability level keeps falling and the release cadence is measured in weeks. OpenAI’s GPT-5.6 Sol, Terra, and Luna entered a gated preview days earlier; Anthropic pushed Claude Sonnet 5 across every plan tier; and Google’s Gemini 3 has continued to press on reasoning and context length. Each release nudges the others, and Grok 4.5’s contribution to that pressure is squarely on cost.

The wedge xAI is choosing is the same one it chose with Grok 4.3: compete on economics. Where 4.3 undercut rivals on sticker price alone, 4.5 adds token efficiency as a second lever — cheaper per token and fewer tokens per task. That is a more durable form of the cost argument, because it survives even when a competitor matches the per-token price.

What is still unverified

A few things will only become clear once independent evaluations arrive:

• Third-party benchmark numbers. SpaceXAI’s own charts are a starting point, not a verdict. The comparisons that matter most will come from evaluations xAI does not control, and the token-efficiency claim in particular deserves independent replication.
• Real-world agentic reliability. A 64.7% score on SWE-Bench Pro is not a 64.7% success rate on a company’s messy internal codebase, where long tool chains expose failure modes clean benchmarks do not.
• The context-window trade-off. Cutting the window from 1M to 500K is an unusual move in a market racing toward longer context. Whether that reflects an efficiency choice or a capability ceiling on the V9 foundation is worth watching.

What it means

Grok 4.5 is best understood not as a bid for the top of the leaderboard but as a bid for the best price-performance ratio in agentic coding. On the benchmarks SpaceXAI published, it is a strong second to Opus 4.8 — ahead on some tests, behind on others — but its token efficiency and aggressive pricing change the calculation for anyone paying per task rather than per benchmark point.

Who benefits. Teams running high-volume coding and agentic workloads get a near-frontier model at a fraction of the effective cost, especially inside Cursor, where the co-training should show up as smoother in-editor behavior. Developers already inside the SpaceXAI ecosystem get a meaningful default upgrade with no migration cost.

The competitive read. By leading with token efficiency and price rather than a clean benchmark sweep, SpaceXAI is telling the market where it believes it can win — the total cost of completing real work, not the top-line accuracy score. That reframes the contest from “who has the smartest model” toward “who has the cheapest reliable one,” a shift that favors whoever can drive down the cost of a completed task fastest.

What to watch. The independent SWE-Bench and DeepSWE numbers will either confirm or puncture the Opus-class framing, and the token-efficiency gap is the single figure most likely to move real adoption. If it holds up outside xAI’s own harness, Grok 4.5 becomes the default choice for cost-sensitive agentic work — and every rival will feel pressure to answer on economics, not just capability.