xAI launched Grok 4.5 on July 8, 2026, and the company is framing it differently than any prior Grok release. Instead of another general chatbot update, xAI built this one for coding and agentic tasks, the kind of multi-step work where a model has to write code, run it, check the output, and fix its own mistakes without a human hovering over every step. The announcement calls Grok 4.5 the company’s “most intelligent offering to date designed for coding and agentic tasks,” and it ships live inside Grok Build, xAI’s coding agent tool, inside Cursor across every plan tier, and through the standard xAI API and console.
The timing is not an accident. OpenAI and Anthropic have spent 2026 locked in a release cadence that barely lets either company breathe, and xAI is now inserting Grok 4.5 directly into that fight with a pricing structure built to undercut both. At $2 per million input tokens and $6 per million output tokens, with cached input priced separately, Grok 4.5 lands well below several of the flagship coding models from its two biggest rivals. Whether cheap tokens translate into developer loyalty is the real question this launch raises, and it is one the market will answer over the next few months, not the next few days.
What xAI actually shipped with Grok 4.5
Grok 4.5 is xAI’s model built around three stated priorities: coding, agentic tasks, and knowledge work. That framing matters because it marks a shift away from Grok’s earlier identity as a conversational assistant tied tightly to X (formerly Twitter) and toward something closer to a developer tool competing head-on with GitHub Copilot’s backing models, OpenAI’s Codex lineage, and Anthropic’s Claude line. According to the announcement, Grok 4.5 carries a model ID of grok-4.5 in the API and supports configurable reasoning effort at three levels: low, medium, and high, with high set as the default.
That reasoning-effort toggle is worth pausing on. It lets a developer trade latency and cost for depth of reasoning on a per-request basis, a pattern that has become standard across the top labs this year. Set it to low for a quick autocomplete-style suggestion, or crank it to high when the model needs to plan out a multi-file refactor. xAI’s own materials describe the product as its model “for coding, agentic tasks, and knowledge work,” language that shows up consistently across the company’s documentation and press materials for this release.
Distribution is the other half of the story. Rather than launching Grok 4.5 as an isolated product, xAI pushed it out through channels developers already use daily. It is available now in Grok Build, the company’s own coding-agent interface. It is also live in Cursor, the AI-native code editor that has become one of the most widely used tools among professional developers switching between models mid-session. And it sits behind the standard xAI console and API, meaning any team already integrated with xAI’s infrastructure can swap in Grok 4.5 without touching their pipeline architecture.
Pricing: the number that will decide adoption
Pricing sits at the center of xAI’s pitch. Grok 4.5 costs $2 per million input tokens and $6 per million output tokens on the standard API, a rate structure disclosed directly in xAI’s own pricing documentation for the release. For teams running high-volume agentic workflows, where a single task can chain together dozens of tool calls and intermediate reasoning steps, token costs compound fast. A model priced meaningfully below the top tier from OpenAI or Anthropic changes the calculus for anyone running agents at scale rather than answering one-off chat queries.
That said, price alone rarely wins a developer market. Coding agents live or die on reliability: does the model actually finish the task, does it hallucinate a nonexistent API, does it get stuck in a loop trying to fix its own broken code. xAI is betting that Grok 4.5’s blend of price and capability clears the bar for a large enough share of coding workloads to matter, even if it isn’t the single top-scoring model on every benchmark. That’s a familiar strategy in this market: undercut on cost, integrate deeply into tools developers already have open, and let volume do the rest.
| Model | Company | Input ($/1M tokens) | Output ($/1M tokens) | Primary use case |
|---|---|---|---|---|
| Grok 4.5 | xAI | $2.00 | $6.00 | Coding, agentic tasks, knowledge work |
| GPT-5.6 Terra | OpenAI | $2.00 | $12.00 | General coding and reasoning |
| GPT-5.6 Luna | OpenAI | $0.20 | $1.20 | Lightweight, high-volume tasks |
| GPT-6 Astra | OpenAI | $10.00 | $50.00 | Flagship reasoning and coding |
| Claude Sonnet 5 | Anthropic | $2.00 | $10.00 | General coding and agents |
| Claude Opus 5 | Anthropic | $5.00 | $25.00 | Complex reasoning tasks |
| Claude Fable/Mythos 5.1 | Anthropic | $10.00 | $50.00 | Long-context, adaptive thinking |
Reading that table, the pattern is clear. Grok 4.5’s output pricing at $6 undercuts every comparable mid-tier and flagship coding model from OpenAI and Anthropic except OpenAI’s stripped-down Luna tier, which is aimed at simple, high-frequency tasks rather than complex agentic coding work. Against the models people actually reach for when running serious coding agents, Grok 4.5 is the cheapest option that still claims flagship-level ambitions.
How Grok 4.5 stacks up on benchmarks
Price is only half the argument. Independent benchmarking from Artificial Analysis, whose Coding Agent Index blends results from DeepSWE, Terminal-Bench v2, and SWE-Atlas Q&A into a single composite score, puts Grok 4.5 running through the Grok Build harness at a score of 76, placing it third overall among evaluated coding agents. That is a genuinely strong result, but it is not first place. OpenAI’s and Anthropic’s most capable coding models still lead on raw benchmark performance, according to the same reporting.
What that gap in absolute performance means in practice depends entirely on the use case. For teams running latency-sensitive, high-stakes production code review, the extra few points of benchmark accuracy from a top-ranked model might be worth the higher token cost. For teams running thousands of parallel coding-agent tasks, testing pull requests, drafting boilerplate, or triaging bug reports, a model that finishes third on capability but costs a fraction as much per task can easily deliver a better return once you multiply out the volume.
xAI has also positioned Grok 4.5’s training as directly tied to real coding workflows, training the model with reinforcement learning against large numbers of multi-step software engineering tasks graded by both automated tooling and model-based evaluation. That training approach lines up with the model’s tight integration into Cursor, a tool built specifically around iterative, agentic code editing rather than single-shot code generation.
The competitive picture: OpenAI, Anthropic, and now xAI
Twelve months ago, the AI coding assistant race was mostly a two-way conversation between OpenAI’s Codex-descended models and Anthropic’s Claude line, with GitHub Copilot serving as the interface most developers actually touched. That framing no longer holds. xAI’s push into coding and agentic tasks with Grok 4.5 adds a legitimate third contender, backed by deep integration into Cursor, one of the fastest-growing code editors among professional developers switching models on the fly.
OpenAI’s response so far has been to keep shipping. The company’s GPT-5.6 family, including the Sol, Terra, and Luna variants, along with a flagship model reported as GPT-6 Astra, gives OpenAI a wide spread of price and capability tiers, from ultra-cheap high-volume options to premium reasoning models priced well above Grok 4.5. Anthropic has taken a similar tiered approach with Claude Haiku 4.5 at the budget end, Claude Sonnet 5 in the middle, and Claude Opus 5 and the Fable/Mythos 5.1 line at the premium end, the latter emphasizing long-context reasoning and what Anthropic calls adaptive thinking.
The strategic split is becoming clearer with each release cycle. OpenAI is betting on ecosystem depth and raw capability at the top end. Anthropic is leaning into long-context reasoning and safety-oriented positioning, charging a premium for its most capable long-context models. xAI is playing a volume game, undercutting on price per token while leaning on Cursor and Grok Build to make sure developers encounter Grok 4.5 inside tools they already use rather than asking them to switch platforms entirely.
| Company | Strategic focus | Distribution channel | Pricing position |
|---|---|---|---|
| xAI | Cost-efficient coding agents | Grok Build, Cursor, xAI API | Aggressive undercut on mid-tier |
| OpenAI | Ecosystem depth, top-end capability | ChatGPT, API, Codex, third-party integrations | Wide tiered spread, premium flagship |
| Anthropic | Long-context reasoning, safety framing | Claude apps, API, enterprise partnerships | Premium on top-tier, moderate on Sonnet |
Why coding and agentic tasks became the battleground
It’s worth stepping back to ask why every major AI lab is racing toward the same target. Coding is one of the few domains where a model’s output is objectively checkable: code either compiles and passes tests, or it doesn’t. That makes it an ideal proving ground for benchmarking claims, and it also happens to be one of the highest-value use cases in the enterprise market. Developer tools carry real willingness to pay, and once a team standardizes its CI pipeline or code review process around a particular model, switching costs go up fast.
Agentic tasks push that further. Rather than answering a single prompt, an agent has to plan a sequence of actions, execute them (often by calling external tools or running code), evaluate the results, and adjust. That loop is exactly what Grok 4.5’s reinforcement learning training against multi-step engineering tasks is designed to strengthen. It’s also exactly the kind of workflow where token costs balloon, since a single agentic run might involve dozens of back-and-forth exchanges rather than one prompt and one response. That’s why pricing per million tokens has become such a visible battleground metric this year: it directly determines whether running an agent fleet at scale is financially sane.
A brief history: how we got from chatbots to coding agents
Grok itself launched in late 2023 as xAI’s answer to ChatGPT, tightly wired into X and marketed initially on personality and real-time data access rather than raw coding capability. Over the following two years, xAI iterated through several major versions, gradually building out reasoning capability and, more recently, acquiring Cursor to gain a direct foothold in the code-editor market rather than competing purely at the model layer.
That acquisition is the missing piece that explains why Grok 4.5 looks so different from earlier Grok releases. Owning the editor that many developers already use daily gives xAI a distribution advantage that pure API access never could. OpenAI took a comparable path by building Codex-branded coding tools directly into its own ecosystem, and Anthropic has leaned on direct API partnerships and enterprise deals with companies embedding Claude into their own developer platforms. Each lab arrived at the same conclusion by a different route: to win the coding market, you need to be where the code gets written, not just answer questions about it after the fact.
What this means for developers choosing a model today
For a developer or engineering team deciding which model to route coding-agent traffic through, the calculus now involves at least three variables: raw capability on your specific task type, cost per task at your expected volume, and how tightly the model integrates with the tools your team already uses. Grok 4.5’s case is strongest on the second and third of those. Its Coding Agent Index score of 76 and third-place ranking suggest it will handle a solid majority of common coding tasks competently, even if it isn’t the top performer on the hardest edge cases.
Teams already living inside Cursor have the lowest-friction path to trying Grok 4.5, since it’s available across every Cursor plan without additional setup. Teams building custom agent infrastructure through the API will need to weigh the $2/$6 pricing against GPT-5.6 Terra’s identical input cost but higher output cost, or against Claude Sonnet 5’s identical input cost paired with a $10 output rate. For pure cost-per-task efficiency at scale, Grok 4.5 currently has a real edge on the models it’s most directly comparable to in capability tier.
Market impact: what changes for OpenAI and Anthropic
Neither OpenAI nor Anthropic is likely to treat Grok 4.5 as an existential threat overnight, but pricing pressure from a well-funded, fast-moving competitor rarely stays contained to just one product line. OpenAI has already shown a willingness to cut prices this year, dropping GPT-5.6 Terra’s rate from $2.50/$15 down to $2/$12 in a July pricing update, well before Grok 4.5 even launched. That kind of proactive repricing suggests OpenAI is already anticipating margin pressure across its mid-tier lineup, and Grok 4.5’s arrival gives it another reason to keep cutting rather than holding steady.
Anthropic’s position is a little different. Its pricing skews toward a premium narrative built around long-context reasoning and adaptive thinking in the Fable/Mythos line, a positioning that doesn’t compete head-on with Grok 4.5’s mid-tier cost play. But Claude Sonnet 5, priced at the same $2 input rate as Grok 4.5 with a higher $10 output rate, sits in more direct competition, and Anthropic may face pressure to adjust that gap if developers start defaulting to Grok 4.5 for cost reasons on comparable workloads.
The broader market impact may be less about any single company losing developers overnight and more about pricing across the industry compressing faster than it otherwise would have. Three well-capitalized labs racing to underprice each other on coding-agent tokens is good news for developers and bad news for anyone counting on high per-token margins from this specific product category holding steady into next year.
Risks and open questions
A few things about Grok 4.5’s launch remain genuinely uncertain. xAI’s own materials and third-party reporting are consistent on pricing, availability, and the model’s stated focus areas, but there’s less independent verification available yet on how it performs on messier, real-world codebases outside curated benchmark suites. Benchmark scores like the Coding Agent Index’s 76 are useful directional signals, not guarantees of how a model behaves on a specific team’s legacy code, unusual frameworks, or domain-specific tooling.
There’s also the question of how sustainable this round of price cuts is for any of the three labs. Running frontier-scale models at $2 input and $6 output per million tokens is only profitable if the underlying compute costs keep falling in step, or if the labs are willing to subsidize adoption in the short term to build market share. Given how much capital-intensive infrastructure all three companies are still building out, it wouldn’t be surprising to see pricing shift again within the next two or three release cycles, in either direction.
Predictions: where this goes next
- Expect OpenAI and Anthropic to respond with further pricing adjustments to their mid-tier coding models within the next one to two quarters, following the pattern OpenAI already set with its July Terra and Luna cuts.
- Cursor’s role as a shared integration point for models from multiple labs will likely deepen, making the editor itself, rather than any single model, the more durable competitive asset in this fight.
- Benchmark suites like the Coding Agent Index will keep gaining visibility as a reference point developers actually cite when choosing a model, similar to how MMLU and other benchmarks became shorthand in the broader LLM race.
- xAI is likely to keep leaning on aggressive reasoning-effort configurability (low/medium/high) as a differentiator, letting developers fine-tune the cost-versus-depth tradeoff on a per-call basis, a feature other labs will likely mirror if it gains traction.
- Expect continued blurring between “chatbot” and “coding agent” product lines across all three companies, as each lab tries to capture both casual users and high-volume developer traffic from a shared model family rather than fully separate products.
Grok 4.5 versus the broader coding-tool market
Model pricing tables only tell part of the story, because most developers don’t call a raw API directly. They work through a layer of tooling: GitHub Copilot, Cursor, Windsurf, JetBrains AI Assistant, and a growing list of agent frameworks that sit on top of whichever underlying model a team chooses. Grok 4.5’s launch strategy leans hard on one of those layers, Cursor, rather than trying to build a rival IDE from scratch. That’s a meaningfully different approach than GitHub Copilot, which is bundled tightly into Microsoft’s own developer ecosystem and largely draws on OpenAI’s models under the hood.
That distinction matters for adoption speed. A developer already paying for Cursor doesn’t need to sign up for a new service, install a new extension, or change their workflow to try Grok 4.5, they just switch a model dropdown. Compare that to the friction of moving an entire team off GitHub Copilot or evaluating a brand-new IDE, and it’s clear why xAI chose to buy its way into an existing editor rather than compete for IDE market share directly. The tradeoff is that xAI now depends partly on Cursor’s continued popularity and neutrality across model providers, a dependency OpenAI and Anthropic don’t share to the same degree since both maintain deeper native integrations elsewhere.
What to watch over the next 90 days
A handful of signals will show whether Grok 4.5’s launch translates into real market share rather than just headlines. Watch whether Cursor’s own usage telemetry or public statements show meaningful uptake of Grok 4.5 relative to OpenAI and Anthropic models already available on the platform. Watch for whether OpenAI or Anthropic announce further price cuts on their mid-tier coding models, which would suggest Grok 4.5’s pricing is being felt internally even without either company naming it directly. And watch independent benchmark trackers like Artificial Analysis for updated Coding Agent Index rankings, since a single launch-week score of 76 is an early data point, not a settled verdict on where Grok 4.5 lands against models that are themselves still being updated.
It’s also worth watching whether xAI follows this release with dedicated enterprise packaging, similar to how OpenAI and Anthropic both offer enterprise-tier agreements with custom rate limits, data handling terms, and support commitments. A strong developer-facing launch is one thing, landing large enterprise coding contracts against incumbents with years of enterprise sales relationships already built is a separate and harder challenge.
Frequently asked questions
When did Grok 4.5 launch?
xAI launched Grok 4.5 on July 8, 2026, positioning it as the company’s most capable model to date for coding and agentic tasks.
How much does Grok 4.5 cost to use?
Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens through the xAI API, with reasoning effort configurable at low, medium, or high (high is the default).
Where can I use Grok 4.5?
Grok 4.5 is available in Grok Build (xAI’s coding agent tool), inside Cursor across all plan tiers, and through the standard xAI console and API using a model ID of grok-4.5.
How does Grok 4.5 compare to GPT-5.6 and Claude Sonnet 5 on price?
Grok 4.5 matches GPT-5.6 Terra and Claude Sonnet 5 on input price at $2 per million tokens, but undercuts both on output pricing at $6 per million tokens, versus $12 for GPT-5.6 Terra and $10 for Claude Sonnet 5.
Is Grok 4.5 the best coding model available right now?
Not by raw benchmark score. On Artificial Analysis’s Coding Agent Index, Grok 4.5 scored 76 and ranked third overall, behind OpenAI’s and Anthropic’s top-tier models on absolute capability, though it remains highly competitive on a cost-per-task basis.
What is Grok Build?
Grok Build is xAI’s own coding agent interface, one of the primary channels through which Grok 4.5 is available at launch, alongside Cursor and the general xAI API.
Does Grok 4.5 support adjustable reasoning depth?
Yes. Developers can set reasoning effort to low, medium, or high depending on whether they need a fast, lightweight response or deeper multi-step reasoning, with high set as the default.
Why is xAI focusing on coding and agentic tasks instead of general chat?
Coding is one of the highest-value, most objectively measurable use cases for AI models, and xAI’s acquisition of Cursor gave the company direct distribution into a developer tool millions of programmers already use, making coding agents a natural focus for Grok 4.5.
Sources: xAI product documentation and announcement materials, xAI developer docs, Artificial Analysis Coding Agent Index, OpenAI model documentation, Anthropic model documentation, Cursor, and background on xAI.
