Quick answer. Grok 4.5 was SpaceXAI's (xAI's) flagship from July 8, 2026 until Grok 4.6 replaced it on August 12, 2026. Musk pitches it as an "Opus-class model, but faster, more token-efficient and lower cost." It ships a 500K-token context window at $2 per million input tokens and $6 per million output tokens (cached input $0.30). Independent testing by Artificial Analysis ranks it #4 of 168 models on its Intelligence Index (score 54) — near the frontier, at a fraction of rivals' cost per task.
🆕 Newer model available. SpaceXAI replaced Grok 4.5 with Grok 4.6 on August 12, 2026 — 35 days later. Same $2/$6 headline pricing and the same 500K context, but a new xhigh reasoning level, a +9-point independently measured SWE-bench Verified jump, and a higher cache price. Read the Grok 4.6 launch guide for confirmed specs, benchmarks, and where it regressed.
SpaceXAI — SpaceX's AI division, still widely referred to as xAI — is moving fast again. On July 8, 2026 it shipped Grok 4.5 to developers through Grok Build, Cursor, and the SpaceXAI API, with the public rollout on grok.com and the X app the next day. Elon Musk's framing was deliberate: an "Opus-class model, but faster, more token-efficient and lower cost." His internal read was more measured — "roughly comparable to Opus 4.7, but much faster."
Unlike the pre-launch guesses, the real numbers are now confirmed. Here's what Grok 4.5 actually is, how much it costs, how it benchmarks against Claude Opus and GPT, and whether it's worth switching to.
What is Grok 4.5?
Grok 4.5 is the successor to Grok 4.3, and it was the newest model in the Grok family until Grok 4.6 shipped on August 12, 2026. It's the first frontier model SpaceXAI has released since acquiring the AI coding startup Cursor — and that acquisition is the real story behind it. According to Cursor, Grok 4.5 is a mixture-of-experts model, jointly trained by Cursor and SpaceXAI on trillions of tokens of real developer-agent data — the debugging traces, tool interactions, and multi-step coding sessions Cursor sees every day.
SpaceXAI has not published an official parameter count or spec sheet; some secondary reports cite a "~1.5 trillion-parameter V9 foundation," but that figure isn't confirmed by the company or by independent trackers, so treat it as unverified. What is confirmed is the positioning: a near-frontier reasoning model tuned hard for agentic coding and knowledge work, sold on price and efficiency rather than raw benchmark supremacy.
One naming note worth getting right: recent launch materials and coverage use SpaceXAI, though "xAI" still appears interchangeably. The model itself is called grok-4.5 in the API.
What did Elon Musk say about Grok 4.5?
Musk's announcement leaned on three claimed advantages:
- Opus-class capability — benchmarked internally against Claude Opus, which SpaceXAI set as its target. His more precise phrasing was that it's "roughly comparable to Opus 4.7," not that it beats Anthropic's current frontier (Opus 4.8).
- Faster — higher throughput than comparable reasoning models (though, with reasoning on by default, time-to-first-token is relatively slow).
- More token-efficient and lower cost — it aims to reach the same answer using far fewer tokens, which is where the real savings show up at scale.
He credited "strong positive feedback from customers in our beta test program" — SpaceX and Tesla engineering teams that ran it privately in late June — as the trigger for the public release. SpaceXAI has also said it plans to ship a new foundation model roughly every month through the end of 2026, an aggressive cadence aimed at closing the gap with Anthropic, OpenAI, and Google.
Grok 4.5 specs and architecture
Here's the confirmed spec sheet, cross-checked against OpenRouter and Artificial Analysis rather than vendor marketing alone:
| Attribute | Grok 4.5 | Grok 4.6 | Grok 4.3 |
|---|---|---|---|
| Maker | SpaceXAI (xAI) | SpaceXAI (xAI) | xAI |
| Architecture | Mixture-of-experts, co-trained with Cursor | Not disclosed — no parameter count, no system card | V8-small |
| Context window | 500K tokens | 500K tokens (unchanged) | 1M tokens |
| Input | Text + image | Text + image | Text, image, video |
| Output | Text only | Text only | Text |
| Reasoning | "High" effort on by default (non-disableable) | low / high (default) / xhigh (new) | Always-on reasoning |
| Measured throughput | ~86.7 tokens/sec (Artificial Analysis, July 2026 harness) | 65.5 tokens/sec (Artificial Analysis, August 2026 harness) | — |
| API model string | grok-4.5 | grok-4.6 | grok-4.3 |
| Status | Superseded by Grok 4.6 on August 12, 2026 | Current flagship | Live |
Two details stand out. First, the context window actually shrank — 500K tokens versus Grok 4.3's 1M — so if you rely on very long contexts, 4.3 still has the edge there. Second, the model supports the full modern tool-use stack: native function calling, structured JSON outputs, web search, X (Twitter) search, and code execution, with "high" reasoning always on. That last point means slower time-to-first-token (Artificial Analysis measured ~16.7 seconds) in exchange for stronger multi-step results. The Cursor co-training is aimed squarely at where Claude Opus has historically been strongest — real agentic coding — which makes 4.5 a genuine option for agentic coding workflows.
How much does Grok 4.5 cost?
SpaceXAI's API pricing is confirmed:
| Model | Input / 1M | Output / 1M | Cached input / 1M |
|---|---|---|---|
| Grok 4.6 (current) | $2.00 | $6.00 | $0.50 (−75%) |
| Grok 4.5 | $2.00 | $6.00 | $0.30 (−85%) |
| Claude Opus 4.8 | $5.00 | $25.00 | — |
| GPT-5.5 / 5.6 | $5.00 | $30.00 | — |
| Claude Fable 5 | $10.00 | $50.00 | — |
| Muse Spark 1.2 | $1.25 | $4.25 | $0.15 |
A high-context surcharge applies above 200K tokens — cross that line and the entire request bills at $4/$12. Grok 4.6 held the $2/$6 headline but raised cached input from $0.30 to $0.50 per million (+67%), so cache-heavy pipelines are one of the few places the older model is genuinely cheaper. Note the nuance: Grok 4.5's output price ($6) is actually higher than Grok 4.3's ($2.50) — the "lower cost" framing is relative to rivals like Opus and GPT, not to its own predecessor. Against those rivals the gap is large: Grok 4.5 undercuts Claude Opus 4.8 by roughly 4× on output and Claude Fable 5 by more than 8×. Competitor prices above are from SpaceXAI's own comparison table, reported by The Decoder. Muse Spark 1.2's line is Meta's published August 2026 pricing, added after this post first ran.
The bigger cost story is token efficiency. SpaceXAI says Grok 4.5 used 4.2× fewer output tokens than Opus 4.8 on SWE-Bench Pro (about 15,954 tokens per task versus 67,020). Fewer tokens per task multiplied by a low per-token price is what actually shows up on the monthly bill — and it's why Artificial Analysis measured a cost of roughly $0.31 per task on its Intelligence Index and about $2.59 per coding task, versus around $11.80 for Claude Fable 5.
How does Grok 4.5 perform on benchmarks?
The most credible independent read comes from Artificial Analysis, which tests models on a common harness. Its verdict: Grok 4.5 scores 54 on the Intelligence Index and ranks #4 of 168 models, behind only Claude Fable 5, GPT-5.5, and Claude Opus 4.8 — a jump of about +16 points over Grok 4.3. On agentic work it's stronger still: a Coding Agent Index of 76 (on par with GPT-5.5's Codex) and the top score on τ³-Banking, a tool-use benchmark, at 33% versus GPT-5.5's 31%. Artificial Analysis places it squarely on the price-performance Pareto frontier, calling it "one of the most cost-efficient models to run for near-frontier intelligence."
SpaceXAI's own launch table leads with coding and agentic benchmarks (not the GPQA/AIME reasoning tests it stayed quiet on), and the honest read is that Grok 4.5 mostly trails Opus 4.8 on raw scores — it wins on cost, not on the leaderboard:
| Benchmark (vendor-reported) | Grok 4.5 | Grok 4.6 | Opus 4.8 | GPT-5.5 | Fable 5 |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 83.3% | not published* | 78.9% | 83.4% | 84.3% |
| SWE-Bench Pro (resolve) | 64.7% | not published | 69.2% | 58.6% | 80.4% |
| DeepSWE 1.1 | 53% | 65.9% | 59% | 67% | 70% |
*One trap worth knowing before you compare Grok 4.5 and Grok 4.6 numbers side by side: SpaceXAI switched harnesses for the 4.6 launch and reported Terminal-Bench v3.0 (26%) instead of the v2.1 figure it used for 4.5 (83.3%). Those are different benchmarks on different scales — a v3.0 number placed next to a v2.1 number is meaningless. Independent v2.1 runs of Grok 4.6 do exist and they disagree with each other by ten points (Artificial Analysis 88.4%, Vals 78.3%), so treat any single Terminal-Bench figure for either model as harness-specific.
SpaceXAI also claims the #1 spot on Harvey's Legal Agent Benchmark. One caveat worth flagging: Artificial Analysis measured a high hallucination rate (roughly 54% on its AA-Omniscience test) even as raw accuracy improved — so for high-stakes factual work, verification still matters. All vendor numbers here are SpaceXAI's own; independent per-benchmark reasoning splits (GPQA, AIME, ARC-AGI) weren't published at launch, so don't trust any specific figure for those yet.
Grok 4.5 in Meta's Muse Spark 1.2 charts
Grok 4.5 also turns up as a comparison point in someone else's launch. When Meta shipped Muse Spark 1.2 and its Muse Code terminal agent on August 5, 2026, Grok 4.5 was one of the models on its benchmark charts — and Meta's figures for it do not match SpaceXAI's own:
| Benchmark | Grok 4.5 (SpaceXAI) | Grok 4.5 (Meta's chart) | Muse Spark 1.2 (Meta) |
|---|---|---|---|
| Terminal-Bench 2.1 | 83.3% | 81.6% | 82.9% |
| DeepSWE 1.1 | 53% | 56.6% | 59.3% |
The gap runs in both directions — Meta scores Grok 4.5 lower than SpaceXAI does on Terminal-Bench and higher on DeepSWE — a useful reminder that a benchmark number is inseparable from the harness that produced it. Both sets are vendor-run. Meta published its results as chart images only, and Muse Spark 1.2 is absent from the verified Terminal-Bench, SWE-bench and Aider leaderboards (as is Claude Opus 5, the model Meta led its chart with).
Two further independent readings have landed since, and both sit below SpaceXAI's own figure. Artificial Analysis measured Grok 4.5 at 81.65% on Terminal-Bench 2.1 in its own harness — within rounding of Meta's 81.6% — and the maintainer-run tbench.ai leaderboard recorded 79.3% ± 1.5% for a Cursor CLI run on July 9, 2026. So two labs converge about 1.7 points under the 83.3% vendor claim, and the official board sits 4.0 points under it. That official entry also carries a −9.0% reward-hacking deduction — roughly ten times the next-largest on a 17-row board — meaning the published 79.3% is already net of it. Grok 4.6 has no tbench submission at all yet, so there is nothing to say about whether that pattern repeats; only that the 83.3% most write-ups quote is the vendor's number, and it is the highest of the four measurements available.
The more direct pressure is on price. Grok 4.5's entire wedge is near-frontier quality at $2/$6; Muse Spark 1.2 lists at $1.25/$4.25 with a 1M-token context window, and Meta also sells a $0.10/$0.20 "contributor" tier in exchange for permission to train on your prompts and completions. Read our Muse Spark complete guide before treating that as a like-for-like cheaper option.
Grok 4.5 vs Claude Opus 4.8 vs GPT-5.6
Grok 4.5 shipped into the same competitive window as OpenAI's GPT-5.6. Here's the practical framing:
- vs Claude Opus 4.8 — Opus stays the reference for coding and long-horizon reasoning quality, and it beats Grok 4.5 on most raw benchmarks. Grok's pitch is "the same class, ~4× cheaper and much faster." For a current head-to-head on coding, see our Grok vs Claude Opus vs Gemini comparison and the Claude Opus complete guide.
- vs GPT-5.6 — OpenAI's ecosystem and tooling breadth remain a key reason teams stay on GPT models. Grok's edge is price, token efficiency, and its tight integration with X and SpaceXAI's own dev stack.
- vs Grok 4.3 — 4.5 is the quality and agentic-tooling step up (MoE, Cursor co-training, +16 Intelligence Index points); 4.3 keeps a longer 1M-token context and a lower output price if that's what your workload needs.
The honest summary: Grok 4.5's claim isn't "smartest model." It's near-frontier intelligence at a fraction of the cost per task, with best-in-class agentic tool-use efficiency. That price-and-efficiency wedge — not benchmark leadership — is the whole story, and it's why analysts think it could pressure Anthropic and OpenAI on margins.
How to access Grok 4.5
Grok 4.5 is live across SpaceXAI's surfaces:
- Grok Build — Grok 4.5 was the default here from July 8 until Grok 4.6 took over on August 12, 2026; 4.5 remains selectable. See our Grok Build, Skills and Connectors guide and the Grok Build CLI walkthrough.
- Cursor — available on all plans, with doubled usage in the first week of launch.
- SpaceXAI API / console — for developers, priced per token as above (
grok-4.5). - grok.com and the X app — for X Premium+ and SuperGrok subscribers.
One important limitation: Grok 4.5 was not available in the EU at launch, with SpaceXAI expecting EU access around mid-July 2026. If you're building for European users, confirm availability before you commit.
Should you switch to Grok 4.5?
If you're running production workloads, the sensible play is to test before you switch. Grok 4.5's cost-and-speed story is genuinely strong, and the Cursor-trained coding focus is aimed at the workloads engineering teams care about most. But it trails Opus 4.8 on raw benchmarks and its hallucination rate is a real caveat, so the "Opus-class" label is best read as "near-frontier at a fraction of the price," not "better than Anthropic."
Practical approach: run your own eval set — your real prompts, your real code — against Grok 4.5 and your current model, and compare quality and total token cost per task. For teams standardizing on agentic coding tools, our AI coding agents guide covers how these models plug into day-to-day workflows.
How does Grok 4.5 compare to Grok 4.6?
SpaceXAI shipped Grok 4.6 on August 12, 2026 — 35 days after 4.5, the fastest turnaround in the family so far. Headline price and context window did not move, so this is a behaviour upgrade rather than a repricing:
| Attribute | Grok 4.5 | Grok 4.6 |
|---|---|---|
| Released | July 8, 2026 | August 12, 2026 |
| API model string | grok-4.5 | grok-4.6 |
| Context window | 500K | 500K (unchanged) |
| Reasoning levels | low / high | low / high / xhigh |
| Price per 1M in / out | $2.00 / $6.00 | $2.00 / $6.00 (unchanged) |
| Cached input per 1M | $0.30 | $0.50 (+67%) |
| AA Intelligence Index | 55.76 | 60.92 (4th overall) |
| Vals SWE-bench Verified | 86.60% | 95.60% |
| LiveBench agentic coding | 56.5 | 54.2 (regression) |
| Median time-to-first-token | 8.7s | 31.2s |
| Measured cost per task (AA) | $0.360 | $0.837 |
| Regions | US at launch, EU expected | us-east-1 and us-west-2 only |
Read it as a real but uneven step. The wins are independently measured: Vals AI ran the full 500-instance SWE-bench Verified set and Grok 4.6 landed at 95.60% against 4.5's 86.60%, and on Artificial Analysis's long-horizon AA-Briefcase suite 4.6 finishes tasks in roughly 53 turns to Claude Opus 5's 103 — the turn-efficiency half of SpaceXAI's cost pitch is the part that survives outside testing intact. The regressions are equally real: LiveBench's contamination-controlled agentic-coding column puts 4.6 below 4.5, time-to-first-token tripled, and Artificial Analysis's measured cost per task rose 2.3× because 4.6 spends roughly 47% more output tokens per job. Per-token price held; cost per finished job did not.
Practically: if your workload is cache-heavy, latency-sensitive, or a short agentic loop 4.5 already handles well, staying on grok-4.5 is a defensible call. For hard single-shot code and long-running agents, 4.6 is the better model.
FAQ
When was Grok 4.5 released?
SpaceXAI launched Grok 4.5 on July 8, 2026 for developers (via Grok Build, Cursor, and the API) and opened public access on grok.com and the X app on July 9, 2026, after a private beta with SpaceX and Tesla teams.
How much does Grok 4.5 cost?
The API is $2.00 per million input tokens and $6.00 per million output tokens, with cached input at $0.30 per million (an 85% discount; Grok 4.6 raised this to $0.50) and a higher-context surcharge above 200K tokens. That's roughly 4× cheaper than Claude Opus 4.8 on output.
How big is Grok 4.5's context window?
500,000 tokens — which is actually smaller than Grok 4.3's 1M-token window. It supports text and image input and produces text output.
Is Grok 4.5 as good as Claude Opus?
On raw benchmarks, mostly no — Grok 4.5 trails Claude Opus 4.8 on tests like SWE-Bench Pro and DeepSWE. Independent testing (Artificial Analysis) ranks it #4 overall, behind Fable 5, GPT-5.5, and Opus 4.8. Its advantage is cost and speed: near-frontier quality at a fraction of the price per task.
What is Grok 4.5 built on?
It's a mixture-of-experts model, jointly trained by Cursor and SpaceXAI on trillions of tokens of developer-agent session data after SpaceXAI acquired Cursor. SpaceXAI hasn't published an official parameter count; the widely-repeated "1.5 trillion parameter V9" figure is unconfirmed.
Is Grok 4.5 available in the EU?
Not at launch. SpaceXAI expected EU availability around mid-July 2026, so European developers should confirm access before building on it.
Is Grok 4.5 the same as Grok 5?
No, and there is still no Grok 5. Grok 4.5 was SpaceXAI's flagship until Grok 4.6 replaced it on August 12, 2026. SpaceXAI has said it plans to ship a new foundation model roughly every month through the end of 2026, and the 35-day gap between 4.5 and 4.6 is so far consistent with that cadence.
Is Grok 4.5 still worth using now that Grok 4.6 is out?
For some workloads, yes. Cached input is cheaper on 4.5 ($0.30 versus $0.50 per million), median time-to-first-token is far lower (8.7s versus 31.2s), and Artificial Analysis measured 4.5 at $0.360 per task against 4.6's $0.837. On raw capability 4.6 is clearly ahead — 95.60% versus 86.60% on Vals' SWE-bench Verified run. Pick 4.5 for latency-sensitive or cache-heavy pipelines, 4.6 for hard code and long-horizon agents.
Building with frontier models and need engineers who already work this way? Codersera helps you hire vetted remote developers who ship with AI coding agents like Grok, Claude, and GPT in the loop.