xAI shipped Grok 4.5 on Wednesday, July 8, 2026, and the number that jumped out first wasn’t a benchmark score. It was the price. At $2 per million input tokens and $6 per million output tokens, xAI priced its new flagship coding and agentic model 76% below Anthropic’s Claude Opus 4.8 on output cost and matched OpenAI’s cheaper GPT-5.6 tier on the output side while undercutting it on positioning. Three weeks on, independent benchmark labs have published enough head-to-head data to answer the question that matters more than the launch-day press release: is Grok 4.5 actually good, or just cheap?
The answer, based on results from Artificial Analysis, Snorkel AI, and a handful of independent testers, is more interesting than either a win or a loss. Grok 4.5 trails Claude Fable 5 and Claude Opus 4.8 on the industry’s standard software-engineering benchmark, wins outright on a marathon-style endurance test, and burns roughly a quarter of the tokens its rivals need to reach a similar answer. That combination is forcing a different kind of comparison than the usual “smartest model” contest, one based on cost per resolved task rather than raw capability alone.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
xAI Launches Grok 4.5 Into a Crowded Coding-Model Race
xAI billed Grok 4.5 as its most capable model to date for coding, agentic workflows, and knowledge work, a narrower pitch than the general-purpose framing xAI used for earlier Grok releases. The company put the model through private testing at SpaceX and Tesla before the public rollout, according to Fullstack Labs’ breakdown of the launch, which lines up with how xAI has folded its consumer AI work more tightly into Elon Musk’s wider group of companies over the past year.
The timing wasn’t an accident. Grok 4.5 landed a little over four months after xAI shipped Grok 4.3, the version tech-insider covered when it launched at a 10x output price gap against Claude Opus 4.8. That gap has now narrowed considerably now that Opus 4.8’s own pricing sits closer to the middle of the field, and Grok 4.5 no longer wins on price by an order of magnitude. It wins by degree, not by a landslide, and that shift says as much about the market maturing as it does about any single model.
Rollout wasn’t instant everywhere. Grok 4.5 stayed unavailable to EU users until mid-July, a delay that has become routine for frontier launches navigating the bloc’s AI Act compliance timelines rather than a Grok-specific holdup.
Inside Grok 4.5: Training, Architecture, and the Cursor Partnership
xAI trained Grok 4.5 on tens of thousands of Nvidia GB300 GPUs running at its Memphis supercluster, the same facility behind earlier Grok generations, now expanded. The reinforcement-learning pipeline covered hundreds of thousands of multi-step software engineering tasks, graded through a mix of automated checks and model-based scoring, a training recipe aimed squarely at agentic coding rather than open-ended chat.
The detail that stands out most is who xAI trained the model alongside: Cursor, the AI-native code editor that has become a default tool for a large share of professional developers. Cursor’s own announcement confirms the model shipped inside the editor on day one, across desktop, web, iOS, and its CLI and SDK, with usage limits doubled for the first week to encourage testing. That kind of co-development deal is new for xAI and signals a shift toward building models around a specific workflow rather than a general chat product retrofitted for coding later.
xAI also pointed to Grok Build, its own agent environment, as the default surface for the model, alongside API access through a console listing the model under the ID grok-4.5.
Grok 4.5 Price: $2/$6 Per Million Tokens, Undercutting Opus 4.8
The standard Grok 4.5 API tier runs $2 per million input tokens and $6 per million output tokens. xAI also offers a faster variant at $4 input and $18 output for workloads that need lower latency, and enabling prompt caching brings the blended effective rate down to roughly $1.35 per million tokens for cache-heavy workloads such as long-running agent sessions.
Here’s how that stacks up against the rest of the field:
| Model | Input ($/M tokens) | Output ($/M tokens) | Notes |
|---|---|---|---|
| GPT-5.6 | $1 | $6 | OpenAI’s current low-cost frontier tier |
| Grok 4.5 (standard) | $2 | $6 | Trained alongside Cursor; GB300-trained |
| Grok 4.5 (faster variant) | $4 | $18 | Lower-latency tier for agent workloads |
| Claude Opus 4.8 | $5 | $25 | Anthropic’s flagship general model |
| Claude Fable 5 | $10 | $50 | Anthropic’s top benchmark scorer |
Two things stand out. First, Grok 4.5’s output price ties GPT-5.6’s exactly, meaning xAI isn’t trying to win the race to the bottom on price alone, it’s positioning against GPT-5.6 on cost while positioning against Opus-class models on capability. Second, the gap to Claude Fable 5 is now the widest in the field: Fable 5’s output rate runs more than 8x Grok 4.5’s. Whether that gap is justified depends entirely on what the benchmarks actually show.
Grok 4.5 Benchmarks: Where It Wins and Where It Doesn’t
xAI didn’t publish a full benchmark sheet at launch, which left independent testers to fill the gap. Results published by webtooltip.com and cross-checked against The Decoder’s independent write-up converge on a consistent picture across four separate test suites.
| Benchmark | Grok 4.5 | Claude Opus 4.8 | Claude Fable 5 | GPT-5.5 |
|---|---|---|---|---|
| SWE-Bench Pro (resolve rate) | 64.7% | 69.2% | 80.4% | 58.6% |
| Terminal Bench 2.1 | 83.3% | 78.9% | 84.3% | 83.4% |
| DeepSWE 1.1 | 53% | 59% | 70% | 67% |
| SWE Marathon | 29.0% | 26.0% | 24.0% | Not disclosed |
| Cost per resolved task (Coding Agent Index) | $2.49 | Not disclosed | $11.80 | $5.07 |
| Artificial Analysis Intelligence Index rank | 4th | 3rd | 1st | 2nd |
On the benchmark that most engineering teams treat as the industry standard, SWE-Bench Pro, Grok 4.5 sits behind both Claude models and ahead of only GPT-5.5. Claude Fable 5 leads by nearly 16 points. That’s a real gap, not noise, and it shows up again on DeepSWE 1.1, where Grok 4.5 trails Fable 5 by 17 points. Ranked purely on the Artificial Analysis Intelligence Index, an aggregate score built from multiple reasoning and coding tests, Grok 4.5 comes in fourth out of the four models compared here.
But two results cut the other way. On Terminal Bench 2.1, which scores multi-step command-line task completion, Grok 4.5 is within a point of both GPT-5.5 and Claude Fable 5. And on SWE Marathon, a newer endurance-style benchmark that runs models through extended, multi-hour coding sessions rather than isolated tasks, Grok 4.5 actually leads the field at 29.0%, ahead of Opus 4.8’s 26.0% and Fable 5’s 24.0%, a result CryptoBriefing highlighted as the most surprising number to come out of the launch window.
Token Efficiency: Grok 4.5’s Real Selling Point
The number that actually explains Grok 4.5’s pricing strategy isn’t a benchmark score, it’s a token count. On SWE-Bench Pro tasks, Grok 4.5 averaged 15,954 output tokens per resolved task. Claude Opus 4.8 averaged 67,020 tokens on the same tasks, over four times as many. That efficiency gap is why a model that loses on raw resolve rate can still win on cost per completed task: Grok 4.5’s $2.49 average cost per task on the Coding Agent Index beats GPT-5.5’s $5.07 by roughly half and undercuts Claude Fable 5’s $11.80 by nearly 80%.
Independent speed testing backs up the efficiency story. xAI’s official throughput figure sits at 80 tokens per second, though one independent tester clocked bursts up to 92.1 tokens per second, with a measured time-to-first-token around 15 seconds. For agent workflows that chain dozens of model calls in sequence, a shorter answer that arrives faster compounds quickly, even when the underlying model is, by benchmark score, not the smartest one available.
This is the argument The Decoder made directly in its own analysis of the launch: the benchmark gap to Fable 5 and GPT-5.5 may simply not matter for teams running high-volume agentic pipelines, where total cost and latency decide the winner more than a 15-point gap on a single leaderboard.
SWE Marathon and Real Professional Work: The Numbers Nobody Expected
The most unexpected data point out of the Grok 4.5 launch window didn’t come from a coding benchmark at all. Snorkel AI ran roughly 2,000 tasks from its GDPval+ dataset, a professional-workplace reasoning test spanning legal, education, healthcare, and financial-analysis tasks graded against rubrics built by domain experts, with Google’s Gemini 3.5 Flash acting as the automated judge.
Grok 4.5 posted a 29% mean pass rate against expert-built criteria, ahead of GPT-5.5’s 22% and Claude Opus 4.8’s 21%. The gap widened in specific domains: Grok 4.5 hit 40% on legal work versus 27-28% for the other two models, 58% on education tasks versus 35-42%, and 35% on healthcare versus 23-25%. GPT-5.5 only led in one category, construction, and Opus 4.8 led among financial-manager tasks.
Snorkel’s own framing of the results is worth sitting with: across every model tested, including the leader, the pass rate against real expert criteria stayed under a third. Even the strongest model in Snorkel’s test cleared fewer than one in three of the standards a human professional would actually apply, according to the company’s own write-up of the results. That’s less a Grok 4.5 story and more a reminder of how far “benchmark-good” still sits from “expert-good” across the entire industry, Grok 4.5 included.
The Hallucination Trade-off Behind the Accuracy Gains
Not every number from the launch window favors Grok 4.5. The Decoder’s testing found that Grok 4.5’s accuracy on its evaluation set rose from 35% to 52% compared with the prior Grok generation, a real gain. But the same testing found the hallucination rate rose alongside it, from 25% up to 54%.
That pairing matters more than either number alone. A model that answers more questions correctly while also fabricating more often isn’t a straightforward upgrade, it’s a different risk profile. For agentic use cases, where a model’s output can trigger a code change, a file edit, or an API call without a human reviewing every step, a higher hallucination rate is the kind of number that shows up in incident reports months later rather than launch-day coverage. Teams evaluating Grok 4.5 for autonomous or lightly supervised agent work should treat this figure as the one that needs independent verification before wide rollout, not the pricing table.
Where You Can Actually Use Grok 4.5 Right Now
Grok 4.5 is live in three places as of this writing: Grok Build, xAI’s own agent environment, where it’s now the default model, inside Cursor, across desktop, web, iOS, CLI, and SDK, and through xAI’s developer API console under the model ID grok-4.5. A typical API request follows the same pattern developers already use for other frontier models, referencing the model ID directly:
curl https://api.x.ai/v1/chat/completions \
-H "Authorization: Bearer $XAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "grok-4.5",
"messages": [{"role": "user", "content": "Refactor this function for readability."}]
}'EU availability lagged the US and rest-of-world rollout by roughly a week, landing in mid-July rather than at the July 8 launch, consistent with the compliance-driven delay pattern that’s become standard for frontier launches into the European market this year. xAI hasn’t announced enterprise-tier packaging separate from the API console pricing, and there’s no confirmed rollout through X’s Premium+ subscription tier as a distinct product bundle.
The Money Behind Grok 4.5: xAI’s $230 Billion Valuation
Grok 4.5 didn’t launch in a vacuum. xAI closed a $20 billion Series E round in January 2026 at a $230 billion valuation, with participation from Nvidia, Cisco Investments, Valor Equity Partners, Fidelity, Qatar Investment Authority, MGX, Baron Capital Group, and Stepstone Group. xAI said the round would fund data center expansion and continued development of the Grok model line, the same GB300 compute buildout that trained Grok 4.5 six months later.
| Round | Series E |
| Amount raised | $20 billion |
| Post-money valuation | $230 billion |
| Closed | January 2026 |
| Key investors | Nvidia, Cisco Investments, Valor Equity Partners, Fidelity, Qatar Investment Authority, MGX, Baron Capital Group, Stepstone Group |
| Stated use of funds | Data center expansion, Grok model development |
| Combined X + Grok monthly active users | ~600 million (self-reported) |
The combined X and Grok monthly active user figure, about 600 million, is a number xAI gave TechCrunch directly rather than a third-party measurement, and the company hasn’t broken it down between the two products. Still, it’s the clearest scale signal xAI has offered publicly, and it puts Grok’s potential distribution advantage in context: a coding-focused model with a direct pipeline into a 600-million-user consumer app is a different competitive threat than a coding model sold purely through an API.
How Grok 4.5 Fits the Mid-2026 AI Model Landscape
Grok 4.5 arrives into a field that’s stopped competing on a single leaderboard. Anthropic is now running two flagship tracks at once, Claude Opus 4.8 as its general-purpose model and Claude Fable 5 as its top benchmark scorer, priced well above both Opus 4.8 and Grok 4.5. OpenAI has split its lineup between GPT-5.5 and the cheaper GPT-5.6 tier, the one Grok 4.5’s output pricing matches exactly. Google’s Gemini line and the open-source contenders, DeepSeek, GLM, and Qwen, are fighting a separate price war that tech-insider covered in detail when DeepSeek V4, GLM-5.2, and Qwen launched within weeks of each other at a 10x spread in price.
What’s changed since Grok 4.3 is that pricing alone no longer separates the field. When Grok 4.3 launched against Gemini 3.1 Pro at a 4x output price gap, xAI’s pitch was straightforward: pay a fraction of the price for a model that’s merely competitive. Grok 4.5 is a harder sell to characterize in one line, because it’s simultaneously the cheapest model in its performance tier and the weakest performer among the models it’s priced closest to. Buyers now have to decide which variable matters more for their workload, and that decision looks different for a startup running thousands of agent calls a day than it does for an enterprise running a handful of high-stakes coding tasks.
Anthropic’s own Claude Sonnet 5, which debuted at half the API cost of its predecessor, shows the same pattern from the other direction: every major lab is now shipping a “good enough, much cheaper” tier alongside its flagship, and Grok 4.5 slots into that category more than it competes directly with Fable 5 or GPT-5.5 at the top of the leaderboard.
From Grok to Grok 4.5: A Short History of xAI’s Model Line
xAI has kept up an unusually fast release cadence since its first Grok model shipped as X’s built-in chatbot. Grok 4.3 arrived earlier in 2026 already priced well below Claude Opus 4.8, and tech-insider’s coverage at the time found it competitive on price but not dominant on capability, a pattern that has essentially repeated with Grok 4.5. What’s different this time is the explicit narrowing of focus: where earlier Grok releases tried to compete as general chat assistants, Grok 4.5 is xAI’s first model built and marketed specifically around coding and agentic work, co-trained with an outside partner, Cursor, rather than developed and shipped in isolation.
That focus shift mirrors what happened across the industry through 2025 and into 2026, as the biggest labs moved from competing on general intelligence claims to competing on specific, measurable workflows: coding, agentic tool use, and now, per Snorkel’s GDPval+ results, professional-domain reasoning. Grok 4.5 is as much a bet on that narrower framing as it is a bet on any specific benchmark number.
What Grok 4.5 Means for Developers and Enterprise Buyers
For individual developers already inside Cursor, Grok 4.5 is close to a no-decision decision: it’s built into the editor, priced competitively, and fast enough that switching costs nothing but a settings change. The doubled usage limits for the first week make it cheap to test against whatever model a team is currently using for day-to-day coding tasks.
The calculus is different for engineering leads deciding what to standardize on for a whole team, and different again for enterprises evaluating agentic deployments where a model acts with minimal human review. The token-efficiency numbers make Grok 4.5 genuinely attractive for high-volume, well-scoped tasks, code review, test generation, routine refactors, where a 20-point benchmark gap matters less than a 4x reduction in tokens burned per task. The hallucination-rate increase argues for the opposite approach on anything higher-stakes: contract analysis, financial calculations, or unsupervised agent chains touching production systems, where Claude Opus 4.8’s higher SWE-Bench Pro score and unreported (and presumably lower) hallucination profile still carry more weight than a lower invoice.
Teams already using a cost-per-task framework to compare coding models now have a genuinely useful new data point to plug in, rather than another leaderboard entry to eyeball.
5 Predictions for the Next Round of the Coding-Model Race
- Anthropic answers on price before it answers on benchmarks. With Grok 4.5 tying GPT-5.6 on output cost, expect Anthropic to introduce a cheaper Opus-tier option rather than race straight to a Fable 6 announcement.
- More editor-specific co-training deals follow the Cursor model. If training alongside a specific IDE produced a benchmark leader on Terminal Bench and SWE Marathon, competing labs will look for their own distribution partners rather than building generically.
- Hallucination rate becomes a standard disclosure, not an afterthought. Once one major outlet puts a hard number on it, as The Decoder did here, expect procurement teams at regulated enterprises to start asking every vendor for the same figure before signing.
- xAI publishes its own benchmark sheet within one to two model cycles. Letting third parties define the narrative around Grok 4.5’s SWE-Bench Pro gap isn’t sustainable if xAI wants enterprise buyers rather than just Cursor’s existing user base.
- Cost-per-resolved-task replaces raw benchmark score as the headline metric. Grok 4.5’s launch is the clearest evidence yet that a model can lose on the leaderboard and still win the purchase order, and expect the next wave of comparison coverage, on tech-insider and elsewhere, to lead with dollar figures instead of percentage points.
Related Coverage
- Grok 4.3 vs Claude Opus 4.8: 10x Output Price Gap [2026]
- Grok 4.3 vs Gemini 3.1 Pro: 4x Output Price Gap [2026]
- Claude Fable 5 vs Opus 4.8: 11-Point SWE-Bench Gap [2026]
- Claude Sonnet 5 Debuts: 57 Score, Half the API Cost [2026]
- DeepSeek V4 vs GLM-5.2 vs Qwen: 10x Price Gap [2026]
- Best AI Model for Coding: DeepSeek Costs 20x Less [2026]
- More AI & Machine Learning coverage
Frequently Asked Questions
What is Grok 4.5?
Grok 4.5 is xAI’s flagship model for coding, agentic tasks, and knowledge work, launched July 8, 2026. It was trained on tens of thousands of Nvidia GB300 GPUs and co-trained alongside the Cursor code editor.
How much does Grok 4.5 cost?
The standard tier costs $2 per million input tokens and $6 per million output tokens. A faster variant costs $4 input and $18 output per million tokens, and prompt caching can bring the blended rate down to roughly $1.35 per million tokens.
Is Grok 4.5 better than Claude Opus 4.8?
Not on the standard SWE-Bench Pro benchmark, where Opus 4.8 scores 69.2% against Grok 4.5’s 64.7%. Grok 4.5 does lead on the SWE Marathon endurance benchmark and uses roughly 4.2 times fewer output tokens per task, which lowers its effective cost per resolved task well below Opus 4.8’s.
Can I use Grok 4.5 in Cursor?
Yes. Cursor shipped Grok 4.5 support on launch day across desktop, web, iOS, CLI, and SDK, with doubled usage limits for the first week after release.
Does Grok 4.5 hallucinate more than earlier Grok models?
Independent testing reported by The Decoder found Grok 4.5’s hallucination rate rose from 25% to 54% alongside an accuracy improvement from 35% to 52% compared with the prior generation. That trade-off is worth verifying independently before deploying the model in unsupervised or high-stakes agent workflows.
Is Grok 4.5 available in the EU?
Grok 4.5 wasn’t available to EU users at the July 8 launch. Access rolled out to the region in mid-July, consistent with compliance-driven delays other frontier labs have faced under the EU AI Act.
What GPUs was Grok 4.5 trained on?
xAI trained Grok 4.5 on tens of thousands of Nvidia GB300 GPUs at its Memphis-based supercluster, funded in part by the company’s $20 billion Series E round closed in January 2026.
How much is xAI worth after this launch?
xAI closed its Series E at a $230 billion valuation in January 2026, six months before Grok 4.5 shipped. The company told TechCrunch it has roughly 600 million combined monthly active users across X and Grok, though it hasn’t broken that figure out by product.
