Grok 4.5: 500K Context, $2/$6 Pricing & API Access (2026)
Grok 4.5 by xAI (July 8, 2026): 500K context, $2 input / $6 output per 1M tokens, AA Intelligence Index 54. See full pricing, benchmarks, and API details.
Grok 4.5 is xAI's cheaper, agentic-coding-focused model, released July 8, 2026, with a 500,000 token context window and an Artificial Analysis Intelligence Index of 54 (rank 8 of 186). It trades context size and published benchmarks for lower cost and same-day availability in Cursor, Grok Build, and five model gateways.
Grok 4.5, released by xAI on July 8, 2026, is a coding-focused model with a 500,000 token context window and an Artificial Analysis Intelligence Index of 54, ranked 8th of 186 tracked models. It costs less than xAI's own Grok 4.20 flagship and ships same-day inside Cursor and Grok Build.
Provider: xAI · Family: Grok 4
Context window: 500,000 tokens
Input modalities: text, image, tool-calls · Output: text, tool-calls
About Grok 4.5
Grok 4.5 is xAI's coding and agentic-work focused model, released July 8, 2026, roughly ten months after Grok 4 and following the February-March 2026 launch of the larger Grok 4.20 flagship. xAI itself completed a corporate merger into SpaceX earlier in 2026, and the combined entity now trades under the ticker SPCX after a June 12, 2026 Nasdaq listing; Grok 4.5 is the first numbered Grok release to ship fully under that structure. Elon Musk described Grok 4.5 publicly as "Opus-class," later clarifying he meant it lands close to Claude Opus 4.7 on general capability while costing and running meaningfully less per token. xAI has not disclosed the model's parameter count or whether it uses a dense or mixture-of-experts backbone, a departure from the detailed architecture claims made around Grok 4.20's four-agent MoE design. On independent benchmarking, Artificial Analysis scores Grok 4.5 at 54 on its Intelligence Index, placing it 8th out of 186 tracked models and above average for its price tier. xAI's own launch materials do not publish GPQA Diamond, AIME, MMLU-Pro, or SWE-bench Verified scores for this specific release, which is a notable gap versus how xAI marketed Grok 4.20 with a full benchmark table. Output speed measures 89.5 tokens per second on Artificial Analysis's tracking, a mid-pack result rather than a speed record. Buyers who need head-to-head reasoning or coding benchmark numbers against GPT-5.5, Claude Opus 4.7, or Gemini 3.1 Pro will not find them published for Grok 4.5 as of this writing. The context window is 500,000 tokens, a reduction from the 1 million tokens Grok 4.3 supported and far below Grok 4.20's 2 million token multi-agent variant. xAI has not published a separate maximum output token limit; requests must fit within the 500K window total. This tradeoff lines up with the model's positioning as a lighter, cheaper coding assistant rather than a long-document or large-codebase specialist, but it is a real regression for teams that had built long-context workflows on the prior Grok 4.3 release. Modalities are narrower than xAI's flagship line: Grok 4.5 accepts text and image input and returns text output only, with no confirmed audio or video input or output. It supports native function calling, web search, X (Twitter) search, and code execution as built-in tools, plus three reasoning depths (low, medium, high, defaulting to high). xAI's documentation recommends setting a prompt cache key and using context compaction for long-running agent loops, both practical nods to the smaller context budget. Pricing runs $2.00 per 1 million input tokens and $6.00 per 1 million output tokens, with cached input tokens billed at $0.50 per 1 million, a 75% discount. That is roughly 60% more expensive on input and output than Grok 4.3's $1.25/$2.50 rates, a price increase that comes alongside the smaller context window. A batch or provisioned-throughput tier is not published for Grok 4.5. On a rough workload of 1 million input tokens and 200,000 output tokens for a daily coding agent, expect close to $3.20 per day at list price before any caching discount. Grok 4.5 is reachable through the xAI API (Responses API and Chat Completions endpoints), the Grok Build terminal coding agent, Cursor across all plans, Microsoft Office add-ins, and model gateways including OpenRouter, Vercel AI Gateway, Cloudflare, Snowflake, and Databricks Mosaic. At launch it was not yet available in the EU API console, with xAI stating regional availability was expected later in the same month. No SDK language list or numeric rate limits were published in the documentation reviewed. The model is closed-weight and proprietary; xAI has not released Grok 4.5's weights, unlike the older Grok 2 checkpoint it published on Hugging Face. No dedicated model card or system card had been published for Grok 4.5 as of this writing, a gap versus Grok 4, Grok 4 Fast, and Grok 4.1, each of which shipped a PDF model card at data.x.ai. Training data cutoff date is not disclosed. This absence of a fresh safety document matters given the Grok product line's recent record: independent researchers and the Anti-Defamation League flagged serious child-safety and antisemitic-content failures in Grok models through January 2026, and buyers evaluating Grok 4.5 for consumer-facing or minor-accessible products should treat the lack of a current system card as an open question rather than an oversight to ignore. Grok 4.5 fits teams that want a cheaper, faster coding assistant inside Cursor or a terminal agent loop and do not need the largest available context window or a fully documented benchmark sheet. Teams doing long-document analysis, large-codebase refactors above 500K tokens, or work that requires a current, transparent safety card are better served by Grok 4.20, Claude Opus 4.7, or GPT-5.5 until xAI publishes more complete documentation for this specific release.
Pricing
$2.00 per 1M input tokens, $6.00 per 1M output tokens. Cached input tokens are $0.50 per 1M, a 75% discount. No batch or provisioned-throughput tier is published for this release.
Key Features
- Native Function Calling and Tool Use: Built-in function calling, web search, X search, and code execution ship as default tools with no extra setup.
- Three Reasoning Depths: Toggle low, medium, or high reasoning per request; defaults to high for maximum accuracy on agentic tasks.
- Cached Input Discount: Cached input tokens cost $0.50 per 1M versus $2.00 standard, a 75% discount for repeat-context agent loops.
- Cursor and Grok Build Integration: Live same-day in Cursor across all plans and in xAI's own Grok Build terminal coding agent.
- Multi-Gateway Availability: Reachable through OpenRouter, Vercel AI Gateway, Cloudflare, Snowflake, and Databricks Mosaic in addition to the direct xAI API.
Pros
- Cheaper than Grok 4.20 and most Opus-class competitors at $2.00 input / $6.00 output per 1M tokens.
- 75% cached-input discount rewards agent loops that reuse system prompts and tool schemas.
- Live at launch across Cursor, Grok Build, Office add-ins, and five model gateways with no waitlist.
Cons
- Context window cut to 500,000 tokens from Grok 4.3's 1 million, a regression for long-document workflows.
- No published GPQA, AIME, MMLU-Pro, or SWE-bench Verified scores at launch, unlike sibling release Grok 4.20.
- No Grok 4.5-specific model card or system card found as of publication; not yet available in the EU API console.
Benchmarks
- artificial analysis intelligence index: 54
- artificial analysis speed tokens per sec: 89.5
Frequently Asked Questions
What is Grok 4.5 and who built it?
Grok 4.5 is a large language model built by xAI, released July 8, 2026, roughly ten months after the original Grok 4 and a few months after the larger Grok 4.20 flagship. xAI itself merged into SpaceX earlier in 2026 and the combined entity now trades as SPCX following a June 2026 Nasdaq listing. Elon Musk described Grok 4.5 as "Opus-class," later clarifying it lands close to Claude Opus 4.7 on general capability while running cheaper and faster. xAI has not disclosed parameter count or whether the model is dense or mixture-of-experts. It was positioned specifically for coding and agentic work, with training that reportedly drew on real Cursor developer session data. The model shipped same-day inside Cursor, Grok Build, and Office add-ins. Pricing starts at $2.00 per 1M input tokens and $6.00 per 1M output tokens. It carries a 500,000 token context window, smaller than its immediate predecessor Grok 4.3.
How much does Grok 4.5 cost per 1M tokens?
Grok 4.5 costs $2.00 per 1 million input tokens and $6.00 per 1 million output tokens through the xAI API. Cached input tokens, meaning tokens reused from a prior prompt via a cache key, cost $0.50 per 1 million, a 75% discount versus fresh input. No batch API or provisioned-throughput discount tier has been published for this specific release. As a worked example, a daily coding agent processing roughly 1 million input tokens and 200,000 output tokens costs about $3.20 per day at list price before any caching savings. A support chat workload of 1,000 turns at 2K input and 500 output tokens per turn runs close to $7.00. This is a price increase versus Grok 4.3, which billed $1.25 input and $2.50 output per 1M tokens, though 4.3 also carried a larger 1M token context window. The model is not open-weight, so there is no self-hosting cost alternative.
What is Grok 4.5's context window and max output?
Grok 4.5 supports a 500,000 token context window, down from the 1 million tokens Grok 4.3 offered and well short of Grok 4.20's 2 million token multi-agent variant. xAI's documentation does not state a separate maximum output token limit; output tokens count against the same 500,000 token budget as input and history. Long-context recall above the 500K mark has not been independently verified in benchmarks reviewed at publication. xAI's own guidance recommends using context compaction, meaning periodically summarizing or trimming older tool-call history, to work within the reduced budget on long-running agent loops. For document handling beyond 500K tokens, or for multi-file codebase analysis, xAI's own Grok 4.20 (2M context) is the better fit within the same vendor lineup. Competing models like Claude Opus 4.7 and GPT-5.5 should be evaluated directly if very large single-request context is a hard requirement.
How does Grok 4.5 compare on benchmarks vs Grok 4.20?
xAI has not published GPQA Diamond, AIME 2025, MMLU-Pro, or SWE-bench Verified scores for Grok 4.5, in contrast to Grok 4.20, which shipped with a full benchmark table showing 88.9% GPQA Diamond and roughly 75% SWE-bench Verified. The only third-party score available for Grok 4.5 is the Artificial Analysis Intelligence Index at 54, ranking it 8th of 186 tracked models, and an output speed of 89.5 tokens per second, both mid-pack relative to frontier models. This gap in published data makes it difficult to state confidently whether Grok 4.5 beats or trails Grok 4.20 on raw capability. xAI's positioning suggests it trades some benchmark headroom for lower cost and latency rather than aiming to top the flagship, similar to how other vendors ship a cheaper coding-tier model alongside a benchmark-leading flagship. The 500K vs 2M context gap between the two models is verified and public even though the reasoning benchmarks are not. Teams that need verified reasoning or coding benchmark numbers before committing should run their own evaluation on representative tasks rather than assume parity with Grok 4.20. The same caution applies when comparing Grok 4.5 to Claude Opus 4.7 and GPT-5.5, both of which publish fuller benchmark suites than xAI has for this release.
Is Grok 4.5 open source or proprietary?
Grok 4.5 is proprietary and closed-weight; xAI has not released its weights or a license for self-hosting. This differs from xAI's older Grok 2 checkpoint, which the company did publish on Hugging Face. Access to Grok 4.5 is API-only, through the xAI direct API (Responses API and Chat Completions endpoints), Grok Build, Cursor, Office add-ins, and gateways including OpenRouter, Vercel AI Gateway, Cloudflare, Snowflake, and Databricks Mosaic. At launch, Grok 4.5 was not yet available in the EU API console, with xAI stating wider regional availability was expected later in the same month. No SDK language list or numeric rate limits were published in the documentation reviewed at time of publication. There are no announced open-weight variants of Grok 4.5.
What modalities does Grok 4.5 support?
Grok 4.5 accepts text and image input and produces text output only; no audio or video input or output has been confirmed for this release. It supports native function calling, meaning it can call external tools and APIs directly, plus built-in web search and X (Twitter) search as first-class tools rather than external integrations. Code execution is available natively as well. The model offers three reasoning depths, low, medium, and high, defaulting to high for maximum accuracy on agentic tasks; lower settings trade some accuracy for speed and cost. Structured output support (strict JSON schema mode) was not explicitly confirmed in the documentation reviewed. Compared to xAI's own Grok Imagine product, which handles text-to-video generation separately, Grok 4.5 itself does not generate images or video directly; any such output would require calling Grok Imagine as an external tool.
Does Grok 4.5 train on user data?
xAI has not published a data retention or training-on-inputs policy specific to Grok 4.5 as of this writing, and no dedicated model card or system card for this release was found, unlike Grok 4, Grok 4 Fast, and Grok 4.1, each of which shipped a PDF model card at data.x.ai. SOC 2, ISO 27001, HIPAA, and GDPR compliance status for this specific model were not confirmed in the sources reviewed. This gap is worth weighing carefully: independent reporting, including a TechCrunch investigation in January 2026, flagged serious child-safety failures across the broader Grok product line. The Anti-Defamation League separately scored Grok poorly on countering antisemitic content in prior testing, an incident that predates Grok 4.5 but is relevant context for the product line's safety track record. xAI's older model cards for Grok 4 and Grok 4.1 did disclose retention windows and RLHF-based alignment methods, but there is no confirmation those same terms carry over unchanged to Grok 4.5. Enterprise buyers with strict data-governance or compliance requirements should request xAI's current data handling terms directly before deploying Grok 4.5 in production, rather than assuming parity with the documented policies of earlier Grok releases.
Who is Grok 4.5 best for and who should avoid it?
Grok 4.5 fits teams that want a cheaper, faster coding assistant inside Cursor or a terminal agent loop like Grok Build, and that do not need the largest available context window. It suits developers already using OpenRouter, Vercel AI Gateway, or Databricks Mosaic, since Grok 4.5 is live on those gateways from launch, and it rewards high-volume agent loops that reuse system prompts through its 75% cached-input discount. Teams doing long-document analysis or large-codebase work above 500,000 tokens should look at Grok 4.20's 2 million token variant instead. Buyers who need publicly verified reasoning or coding benchmark scores for procurement should be cautious, since xAI has not published a GPQA, AIME, MMLU-Pro, or SWE-bench Verified table for this release. Voice-first or video-native product teams should also avoid it, since Grok 4.5 outputs text only. Organizations with strict safety or compliance sign-off requirements should wait for a dedicated model card before committing.