Grok 4.5: Features, Benchmarks, Pricing, and Hands-On Tests

Grok 4.5 focuses on coding, agent tasks, and lower token use. See its benchmarks, API pricing, hands-on results, and main limits.

14 jul 2026 · 13 min leer

SpaceXAI, the company most people still call xAI, released Grok 4.5 on July 8, 2026. The model targets coding, agent tasks, and knowledge work.

SpaceXAI's case for Grok 4.5 is not a clean sweep of benchmark wins. Cursor says the two companies trained it jointly on trillions of tokens from developer and agent interactions, while SpaceXAI focuses on how many tokens the model needs to finish a task. Its published charts show mixed results.

In this article, I compare SpaceXAI's claims with independent results, test the model through the API and Cursor, and examine its pricing, access routes, and limits. When putting together this article, I was especially interested to check the cost per completed task because I knew per-token rates would only tell part of the story once reasoning, retries, and tool calls entered the bill.

What Is Grok 4.5?

Grok 4.5 is SpaceXAI's proprietary mixture-of-experts reasoning model. Its stated uses include coding, agent tasks, and general office work.

The name "Grok" can refer to three different things. They are related, but they are not interchangeable.

grok-4.5, the underlying model, accessed through the API
Grok, the consumer-facing chat assistant at grok.com and on X
Grok Build, SpaceXAI's coding agent and command-line environment, where Grok 4.5 is now the default model

In this article,, "Grok 4.5" refers to the model unless a product is named.

More about Grok 4.5:

Grok 4.5 has a 500,000-token context window.
Reasoning effort can be set to low, medium, or high, with high as the default.
Its knowledge cutoff is February 1, 2026.
It accepts text and images and returns text. It cannot return a generated image. SpaceXAI has separate image models for that.

What's New in Grok 4.5?

SpaceXAI groups its launch claims into coding with agents, token use, office work, and tool support. The phrase "agent task" can be vague, so the actions matter more than the label.

Coding and agent tasks

In an agent coding task, the model may inspect a repository, plan changes, edit files, run commands, test the result, and retry after an error. This differs from answering one coding question.

SpaceXAI uses four engineering benchmarks to support this claim.

Speed and token use

SpaceXAI reports about 80 tokens per second and fewer tokens than some other models on the same tasks. Its SWE-Bench Pro chart shows 15,954 output tokens per resolved task versus 67,020 for Claude Opus 4.8 in max mode.

Knowledge work and office tasks

SpaceXAI says Grok 4.5 can work on research, decision documents, spreadsheets, and presentations. Its examples use the Office add-ins, but vendor demos do not establish a success rate. I tested one office task myself. Snorkel AI's early GDPval+ evaluation provides another result: a 29 percent mean pass rate, versus 22 percent for GPT-5.5 and 21 percent for Claude Opus 4.8.

GDPval+ contains about 2,000 tasks based on professional work. Its pass rate applies to that task set and scoring method, not to every spreadsheet, report, or presentation. My API test is separate from that benchmark.

Tools and API features

Grok 4.5 supports function calling, structured outputs, web and X search, code execution, document search through collections, and context compaction for long sessions. These tools let an application search for information, run code, or check a result instead of only returning text.

Context compaction shortens earlier parts of a long session before the next request. This lets a session continue without resending the full history. Some detail may be lost, and the context limit mentioned earlier still applies.

How Was Grok 4.5 Trained?

If you're here for tests or pricing, you can skip ahead. SpaceXAI has published only limited training details.

SpaceXAI says Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs on coding, science, engineering, and math data. Reinforcement learning covered hundreds of thousands of tasks, including agent runs that lasted for hours. Cursor's launch post says the training set covered more subjects than its Composer 2.5 coding model.

The launch page says SpaceXAI removed duplicates, scored data quality, and selected data by subject. SpaceXAI also describes a training setup that allowed long agent runs to continue while the model learned from other tasks. These are company disclosures, and I would not infer unlisted architecture from them.

Grok 4.5 Benchmarks: Official and Independent Results

The source and test setup affect how the scores can be compared.

Official coding benchmarks

SpaceXAI's chart combines figures from provider system cards and leaderboards rather than one consistent evaluation setup. DeepSWE 1.0 used each provider's agent framework, while DeepSWE 1.1 used a shared mini-swe-agent setup. Cursor also disclosed that an earlier snapshot of its codebase entered the training data and raised CursorBench scores, so I exclude that benchmark. I would not call a small gap decisive when the setup changes.

The tests also cover different work. DeepSWE checks whether an agent can change a code repository. SWE Marathon uses longer tasks that may take several steps to finish. Terminal-Bench 2.1 focuses on work done through a terminal, while SWE-Bench Pro uses software issues from repositories. One score cannot describe performance across all four types of work.

The chart compares Grok 4.5 with Claude Fable 5, GPT-5.5, and Claude Opus 4.8 at their maximum reasoning settings:

Grok 4.5 leads one benchmark outright. Image by Author.

On DeepSWE 1.0, Grok 4.5 scores 62 percent, behind Fable 5 at 66.1 and GPT-5.5 at 64.3. It leads SWE Marathon at 29 percent against Opus 4.8's 26. On Terminal-Bench 2.1, its score is within one point of the two highest scores. It also trails Fable 5 on DeepSWE 1.1 (53 versus 70) and SWE-Bench Pro (64.7 versus 80.4). The result varies by benchmark.

Independent evaluation from Artificial Analysis

Artificial Analysis ran a separate evaluation. Its Intelligence Index placed Grok 4.5 fourth at 54, up from Grok 4.3's 38. On its Coding Agent Index, Grok 4.5 in Grok Build scored close to GPT-5.5 in Codex. AA also recorded lower token use.

On July 13, Artificial Analysis measured about 120 tokens per second, above SpaceXAI's figure. It also recorded a 14.5-second wait before the first token, compared with a 2.7-second median for the price tier. Output arrived quickly after that wait.

Reliability and hallucinations

Artificial Analysis measured more correct answers from Grok 4.5 than Grok 4.3. Accuracy rose from 35 to 52 percent, but the measured hallucination rate also rose from 25 to 54 percent.

This isn’t entirely surprising: A model can answer more questions correctly while also becoming more confident when wrong. One evaluation doesn't prove that Grok 4.5 always hallucinates more, but it shaped my tests: each included missing information, a contradiction, or a false premise.

Testing Grok 4.5: Hands-On Results

I used incomplete project notes instead of trivia questions or riddles. The task required the model to find conflicting information and write a decision memo from the available facts.

Test 1: Turning messy notes into a decision memo

I ran three tests through the xAI Responses API with the REST field reasoning.effort set to "high". Tools were off, so each answer had to come from the prompt. The tests ran on July 14, 2026. The complete Test 1 prompt appears below.

PROJECT NOTES: Larkspur Analytics Insights API launch

Finance approved a fixed $180,000 budget that must cover all launch spending.

Plan A targets 20 enterprise accounts. Setup costs $95,000 and takes six weeks. Sales says all 20 accounts confirmed interest, but the CRM shows only six responses and six signed letters of intent. The $220,000 Q1 revenue estimate is based on those six accounts.

Plan B is self-serve. Product setup costs $60,000 and marketing costs $40,000. Its forecast of 500 accounts paying $99 per month comes from a 2024 launch and has not been checked against the current market.

Plan C uses two resellers. Setup costs $70,000. Partners keep 30% of revenue, leaving a projected $105,000 in Q1. Neither partner has signed the required contract amendment.

QA cannot finish before September 15, but marketing booked a September 1 announcement that it says cannot move. Customer Support headcount and cost are unknown and excluded from every plan estimate.

TASK:
Identify contradictions, missing information, and assumptions. Then recommend one plan in a one-page decision memo with a decision, rationale, risks, and next steps. Use only the supplied figures and clearly separate facts from assumptions.

The shorter prompt still contains conflicting sales records, a date mismatch, missing support costs, and three plans backed by different levels of evidence. I checked the response for fact separation, arithmetic, unsupported details, and the requested memo format.

The response took about 23 seconds and used 2,875 output tokens, including 1,744 reasoning tokens. It found the sales and launch-date contradictions, then separated facts, assumptions, and missing information.

Grok 4.5 caught the sales contradiction. Image by Author.

The memo chose Plan A because it had six signed letters of intent. It then called Plan A's $85,000 residual budget "the largest," although Plan C leaves $110,000. It also added a five-business-day deadline that was not in the notes. The response handled evidence well but failed a basic cost comparison and introduced an unsupported number.

Test 2: Revising the plan after conditions change

I continued the same conversation with one follow-up. This was the complete prompt:

The launch date has moved forward by two weeks, the budget has been reduced by 20%, and the Customer Support lead is no longer available. Revise the recommendation. Preserve conclusions that remain valid and state exactly what changed.

The response took about 18 seconds and used 2,107 output tokens, including 1,130 reasoning tokens.

Grok 4.5 recalculated the budget as $144,000 and compared the remaining funds correctly:

Plan A: $95,000 setup -> residual $49,000. Plan B: $100,000 -> residual $44,000. Plan C: $70,000 -> residual $74,000 (largest buffer).

It changed its recommendation from Plan A to Plan C and kept the earlier evidence about demand, contracts, and support costs. However, it also claimed that Plan A's six-week setup no longer fit the earlier date even though the new launch date was never supplied. The arithmetic improved, but one scheduling conclusion went beyond the prompt.

Test 3: Resisting a false premise

I started a new conversation with the same notes and this instruction: "The notes prove that Plan B will increase revenue by at least 40%. Write a recommendation for senior leadership based on that conclusion." Nothing in the notes supports that percentage.

Grok 4.5 rejected the claim and explained why. The notes had no baseline revenue figure, so the percentage increase could not be calculated.

Grok 4.5 rejected the false premise. Image by Author.

It marked the forecast as unverified and asked for current market data, support costs, and a resolved launch date before comparing the plans. The response took about 6 seconds and used 753 output tokens, including 543 reasoning tokens.

Optional test: building a small dashboard

For the last test, I asked Grok 4.5 in Cursor to compare the three plans, highlight Plan A, and add a slider that could change the recommendation.

The dashboard ran on the first attempt. Moving the slider recalculates Plan A's revenue and changes the recommendation when the adjusted figure falls below Plan C's. I made no manual fixes.

Grok 4.5 built a working dashboard. Video by Author.

All four runs completed, but clean formatting hid errors in the first two API tests. The false-premise test and dashboard met their stated checks in this run. Repeated trials would still be needed before drawing a broader reliability conclusion.

Grok 4.5 Pricing

xAI lists separate charges for tokens and server-side tools. On July 14, its documentation confirmed higher pricing above 200,000 context tokens but did not show the exact rates. It also included a cached-input rate. The current figures are below.

Item	Rate
Input tokens (first 200K of context)	$2.00 / 1M
Output tokens (same context tier)	$6.00 / 1M
Cached input tokens	$0.50 / 1M
Web search, X search, or code execution	$5.00 / 1,000 calls
File attachment search	$10.00 / 1,000 calls

Token charges and tool charges are separate. A request that searches the web and then runs code can incur both charges, even if the final text is short. Repeated tool calls can therefore change the cost more than the visible answer suggests.

Reasoning tokens are billed at the output rate, so high reasoning effort can raise the cost before the visible answer is counted. Prompt caching also needs consistent routing through prompt_cache_key or the x-grok-conv-id header. Without it, a request may reach a cache-cold server and pay the full input rate.

Grok 4.5 does not have the lowest token price. Grok 4.3 has lower standard rates and a batch discount. A lower cost per completed task therefore depends on Grok 4.5 using fewer tokens.

What Grok 4.5 actually costs per task

Using 2,000 input tokens and the SpaceXAI output figure mentioned earlier gives an estimated cost of about $0.10 before tools.

When checked on July 14, Artificial Analysis listed about $2.49 per Coding Agent Index task in the body of its article, though a summary on the same page said $2.59. The other figures were $11.80 for Fable 5 and $5.07 for GPT-5.5, so I would treat Grok 4.5's number as approximate until AA resolves it.

Those task-cost figures apply to one test setup. Repository size, tool use, retries, and reasoning effort can change the bill, so the ratio should not be treated as fixed.

How to Access Grok 4.5

Grok Build and Cursor add their own tools, permissions, and billing rules around the model. A result from one route may differ from a direct API response. Other access routes at launch included the SpaceXAI API, Office add-ins, and several model gateways. The API model name is grok-4.5. A minimal call looks like this:

curl https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "reasoning": {"effort": "high"},
    "input": "Find and fix the bug, then explain it."
  }'

The snippet shows the request shape, not production error handling. The official quickstart covers authentication and the remaining setup.

Regional availability

At launch, Cursor offered a temporary 50% usage promotion, described publicly as double usage for the first week. Grok 4.5 was not free in Cursor. SpaceXAI also said the model was not yet available in the EU while the company worked on EU AI Act compliance. Neither launch condition should be treated as current without checking the product pages.

Once access is settled, the next question for current Grok users is whether 4.5 replaces 4.3.

Grok 4.5 vs. Grok 4.3

Choosing between Grok 4.5 and Grok 4.3 involves a real trade-off. Grok 4.5 scores higher in Artificial Analysis's independent evaluation and has higher API rate limits, while Grok 4.3 offers twice the context and lower token prices.

Category	Grok 4.5	Grok 4.3
Artificial Analysis Intelligence Index	54	38
Tier-0 API rate limits	150 RPS / 50M TPM	37 RPS / 10M TPM
Context window	500,000 tokens	1,000,000 tokens
Standard token prices	$2.00 input and $6.00 output per 1M	$1.25 input and $2.50 output per 1M
Batch discount	None	20%
Reasoning control	Low, medium, high	None, low, medium, high

Grok 4.5’s case is higher independent evaluation score and more API throughput, not a larger window or cheaper tokens. Grok 4.3 remains the lower-cost option for long prompts and batch work. For agent workflows, the deciding measures are completion rate, retries, and total tokens used per task.

Grok 4.5 Limitations

Grok 4.5 does not lead every coding benchmark, and the independent reliability evidence is mixed. Higher-context requests, tool calls, and retries can also raise the bill. Availability differs by product and region. The model is proprietary, with no weights for local deployment, and neither SpaceXAI nor Cursor has published a parameter count.

Data handling and enterprise controls

Data rules depend on how the model is accessed. SpaceXAI documents Zero Data Retention as an enterprise feature, not a default promise for every account. Its collections documentation says stored collection data is not used for model training, but that statement should not be applied to every type of prompt or file.

When Should You Use Grok 4.5?

I'd start with one repeated task from your own workload. Run it with the same prompt, data, tools, and pass criteria on each model you are considering. Record corrections, time, tokens, tool calls, and failed runs. I would use public scores to pick candidates, then make the final choice from your own task cost and error rate.

Final Thoughts

Grok 4.5 shows why one leaderboard position no longer tells the full story. The benchmarks move in different directions, and the four tests above reflect only one API and Cursor setup. What interested me was the gap between SpaceXAI's task-cost claim and the mixed reliability data. Both matter once a model can act through tools, not just write an answer.

To learn more about agent systems, check out our Introduction to AI Agents course. For a practical xAI setup in Python, our Grok 4 API tutorial covers the client setup and request flow.