Grok 4.5's 16-Point Leap Makes xAI A Frontier Supplier

TL;DR: Grok 4.5 jumped 16 points, from 38 to 54, on the Artificial Analysis Intelligence Index. In the July 10 homepage highlights, it appears behind the 60-point Fable routed configuration and 59-point GPT-5.6 Sol while delivering 90 output tokens per second at a measured $0.31 per Intelligence Index task.^[3][5] The real story isn't that Grok won every ranking. It is that xAI combined frontier-class intelligence, Cursor-trained agent behavior, live search, and aggressive economics in one credible release.

Sixteen points is not an upgrade. It is an arrival.

Grok 4.3 sat at 38 on the Artificial Analysis Intelligence Index. Grok 4.5 scores 54. In one release, xAI moved from the lower edge of the displayed frontier field to within six points of its leader. It did so while Artificial Analysis measured 89.5 output tokens per second and $0.31 per benchmark task.^[3][5] Those three numbers, 54, 90, and $0.31, explain the launch better than another Grok-versus-GPT headline ever could.

The field around it is already too broad for a two-company contest. Artificial Analysis' July 10 homepage highlights show the routed Fable 5 configuration at 60, GPT-5.6 Sol at 59, Grok at 54, GLM-5.2 at 51, and Gemini 3.5 Flash at 50.^[5] The open-model cluster makes the picture more complicated: MiniMax-M3, DeepSeek V4 Pro, and Kimi K2.6 each score 44, but arrive with radically different speed, price, context, and deployment trade-offs.^[10][11][12]

That is why Grok 4.5 matters. It does not dominate every axis. It compresses enough of them into one product to make xAI strategically credible. The model is jointly shaped by Cursor workflow traces, shipped inside Grok Build, exposed through a Responses API, connected to live X and web search, and priced at $2 per million input tokens and $6 per million output tokens.^[1][2]

The Leap: Sixteen Points Changed xAI's Position

xAI, now branded as SpaceXAI on its launch materials, calls Grok 4.5 its smartest model for coding, agentic tasks, and knowledge work. It is a mixture-of-experts model trained jointly with Cursor across coding, science, engineering, mathematics, and long-running tool workflows.^[1][2]

The release changes xAI's position because the gain is broad enough to survive independent measurement. Artificial Analysis recorded a 16-point improvement over Grok 4.3, placing Grok 4.5 fourth on the full index at the time of its July 8 evaluation.^[3] The live rankings will move as new models and configurations arrive. The stable fact is the score: 54, not the temporary label beside it.

Grok 4.5 also turns xAI's model into a distribution event. It became the default in Grok Build, launched across Cursor desktop, web, iOS, CLI, and SDK, and reached the xAI Responses API as grok-4.5.^[1][2][15] The model supports text and image input, configurable reasoning, function calling, structured outputs, and a 500,000-token context window.^[4]

What's often overlooked is that xAI did not need first place to change its category. Before Grok 4.5, the company could still be treated as an unusually well-funded distribution platform attached to X. A 54-point independent score forces buyers to treat it as a model supplier.

The Market Map: Grok's Strength Is Balance, Not A Crown

Artificial Analysis' homepage highlights are a market map, not the complete leaderboard. They compare selected configurations and mix high, max, and adaptive reasoning settings. Fable's 60-point result is specifically Claude Fable 5 with adaptive reasoning, max effort, and Opus 4.8 fallback.^[6] Grok's result uses high; Sol and GLM use max. Those qualifiers matter.

Fable is the routed fallback configuration. Speed is output generation rate after the first output token, not end-to-end response time. Cost per task is Artificial Analysis' weighted benchmark cost, not a universal production price. N/A means the value was absent from the captured homepage highlight panel.^[5]

The chart punctures every easy victory claim. Fable and Sol remain more intelligent in this snapshot. GLM-5.2 and Gemini 3.5 Flash are much faster at 191 and 161 tokens per second. gpt-oss-120b reaches 294 tokens per second. DeepSeek V4 Pro costs just $0.04 per measured task, while MiniMax-M3 costs $0.12.^[5]

Grok's position is the combination. It delivers 90% of the routed Fable configuration's rounded Intelligence score at 11.3% of its measured task cost. Against Sol, it delivers 91.5% of the score at 29.8% of the cost. Those ratios describe this benchmark snapshot, not every workload, but they show why Grok sits on the capability-cost frontier.^[3][6][7]

The uncomfortable truth is that the frontier no longer has one winner. Fable owns the displayed intelligence crown. gpt-oss owns throughput. DeepSeek owns task-cost efficiency. GLM-5.2 leads the open-weight intelligence class at 51 while serving at 191 tokens per second.^[8] Grok owns a commercially useful middle.

The Coding Test: Grok Joins The Lead Pack Without Winning It

The market-wide frame does not excuse weak coding evidence. Grok 4.5 is sold as a coding and agent model, so the harder question is whether it can operate inside real tools, repositories, and terminals.

xAI reports 83.3% on Terminal-Bench 2.1, 64.7% on SWE-Bench Pro, 62.0% on DeepSWE 1.0, and 53% on DeepSWE 1.1.^[1] Its own chart shows Fable leading all four comparisons. On DeepSWE 1.1, Fable reaches 70%, GPT-5.5 67%, Opus 4.8 59%, and Grok 53%. Grok's launch is credible, but it is not a coding sweep.

Artificial Analysis' model-level Coding Index makes the hierarchy less flattering, and more useful. In the July 10 captured view, Grok 4.5 at high reasoning scores 72.4, sixth among the 27 visible configurations. Sol leads at 77.4, Terra scores 76.7, and the routed Fable configuration scores 76.5. Grok sits five points off the lead, but remains ahead of Sonnet 5, Luna, Gemini 3.5 Flash, and GLM-5.2 in this snapshot.^[5]

Fable is the routed fallback configuration. The screenshot's selector says 28 of 570 models, but only 27 bars are visible, so this article does not infer an overall rank across 570 models. The model-level Coding Index equally averages Terminal-Bench v2.1 and SciCode. Grok 4.5's 72.4 is 29.2 points above Grok 4.3's 42.2 in the same captured view.^[9]

Do not merge that table with the Coding Agent Index. The Coding Index measures a selected model configuration across Terminal-Bench v2.1 and SciCode. The Coding Agent Index measures a deployed model-plus-agent system, where the harness, tools, and execution loop affect the result. Artificial Analysis gives Grok 4.5 inside Grok Build a separate 76 on that agent-system index. At the time of its evaluation, that matched GPT-5.5 in Codex and trailed Fable in Claude Code.^[3] GPT-5.6 Sol later raised the Coding Agent Index ceiling to 80, with Terra at 77 and Luna at 75.^[7] The two scores answer different procurement questions. They are not numerically comparable.

Cursor helps explain the jump. The company says Grok 4.5 was trained on trillions of tokens representing user interactions with codebases and software tools. Reinforcement-learning environments taught the model to investigate, call tools, recover from errors, and verify results across software engineering and broader knowledge work.^[2]

That is not just more code. It is process data. But the partnership also produced a warning: Cursor disclosed that an older snapshot of its codebase was accidentally included in training, giving Grok an advantage on CursorBench. Cursor removed the data from future models and excluded the contaminated comparison from launch materials.^[2]

Grok's strongest coding argument is therefore not supremacy. It is that near-frontier agent performance may arrive with fewer tokens and lower task cost. That is a business claim developers can test.

The Product Bet: xAI Is Selling The Agent Loop

Here's the genius: Grok 4.5 is not being shipped as a model file behind an endpoint. It is the default intelligence inside Grok Build, an agent available through an interactive terminal, headless scripts, bots, and the Agent Client Protocol.^[15] The same model is available in Cursor and through the xAI API, so the launch already spans a first-party agent, a major coding environment, and an OpenAI-compatible developer surface.

That distribution can become a feedback loop. Better workflow traces improve the model. A better model attracts more agent work. More work creates more traces. While competitors optimize isolated model endpoints, xAI and Cursor can optimize the interaction between model, harness, repository, and developer.

The other differentiator is live retrieval. Grok 4.5 supports web search, X search, code execution, function calling, and structured outputs.^[4][14] X search can perform keyword, semantic, user, and thread search directly against the platform. xAI charges $5 per 1,000 calls for web search, X search, and code execution, on top of token costs.^[13]

What's often overlooked is that this is not merely a feature list. It is xAI's attempt to bundle a model, agent harness, real-time data source, and distribution channel. Fable may reason better. Sol may score higher. GLM may serve faster. DeepSeek may cost less. Grok's product bet is that fewer buyers want to assemble those pieces themselves.

The Hidden Bill: Verification Can Erase The Discount

Grok 4.5's measured cost advantage is real inside the Artificial Analysis methodology. It is not the total cost of deploying an autonomous agent.

Artificial Analysis found that Grok 4.5's AA-Omniscience accuracy rose from 35% to 52%, while its hallucination rate rose from 25% to 54%.^[3] The model knows more and is more willing to answer. Those are not the same as being more reliable.

The context window has a similar catch. Grok 4.5 supports 500,000 tokens, half Grok 4.3's 1 million.^[3][4] Artificial Analysis reports that pricing doubles once inputs exceed 200,000 tokens.^[3] A large window is a capability ceiling, not a promise that filling it is economical.

Live X retrieval increases the same tension. It gives Grok privileged access to breaking conversations and public sentiment, but also exposes the agent to rumors, coordinated campaigns, impersonation, and confident nonsense at platform speed. Retrieval is not verification.

Cybersecurity is the final unresolved bill. Cursor says Grok 4.5's stronger coding and tool use required new safeguards. It says high-risk workflows are detected and blocked without silently falling back to a weaker model.^[2] That is useful, but xAI did not publish a launch safety report comparable in detail to the most transparent frontier deployments.

The Verdict: Grok 4.5 Makes xAI A Frontier Supplier

Grok 4.5 does not own the intelligence crown, the speed crown, the task-cost crown, or the coding crown. The routed Fable configuration and Sol score higher. gpt-oss is faster. DeepSeek is cheaper per measured task. Fable leads the coding comparisons xAI published. That list does not weaken the release. It explains why the release is strategically important.

The frontier is no longer one ladder. It is a market where buyers trade intelligence, throughput, task cost, agent reliability, context, retrieval, and governance against one another. Grok 4.5 enters that market with a coherent position: frontier-class intelligence, competitive serving speed, aggressive pricing, Cursor-shaped agent behavior, and live information tools in one package.

The real story isn't that Grok beat every rival. It is that xAI improved by 16 points and removed the easiest reason to ignore it. Before Grok 4.5, xAI could be treated as a distribution phenomenon attached to X. After Grok 4.5, it has to be evaluated as a frontier model supplier.

Fable and Sol still own the displayed crown. Grok 4.5's achievement is making the space beneath it commercially dangerous.

Last updated: July 10, 2026