A fresh wave of coverage from outlets including 36Kr, Mashable, NDTV and TheStreet is circulating a single framing of Elon Musk’s current ambitions: that what he is chasing is “not parameters, but unprecedented possibilities humanity has never achieved.” The reporting ties that idea to Tesla’s flying car reveal and a new supercar, both pitched less as products and more as declarations about where technology’s edges actually sit. But the same philosophy is playing out, in far more concrete and measurable terms, inside Musk’s AI company xAI and its push into coding tools — an area where the “parameters versus possibilities” framing has already produced a real, dated, and priced product shift.
Over the past 14 months, xAI has launched an agentic coding model, published benchmark numbers that an independent evaluator could not reproduce, quietly retired that same model, replaced it with a successor, and then turned its consumer-facing Grok Bot into something closer to a dispatcher than a single brain — one that picks among several AI systems depending on the task rather than leaning on raw scale. For developers and engineering leaders watching the AI coding tools market, that sequence is the clearest evidence yet of Musk’s stated priority: capability over parameter count.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Grok Code Fast: xAI’s First Bet on Coding Over Scale
xAI’s entry into the dedicated coding-agent market came on August 28, 2025, with the release of Grok Code Fast 1, a model built specifically for software-development workflows rather than general chat. The pitch was speed and integration rather than sheer size. Grok Code Fast 1 shipped with a 256,000-token context window, large enough to hold substantial codebases or long task histories in a single pass, and xAI advertised a generation speed of roughly 92 tokens per second — fast enough to feel responsive inside an IDE rather than a batch job.
Pricing undercut much of the agentic-coding field at launch: $0.20 per million input tokens and $1.50 per million output tokens, a structure clearly built to encourage high-volume, iterative use rather than occasional one-off queries. xAI also moved quickly to get the model embedded where developers already work, striking integrations with GitHub Copilot, Cursor, and Windsurf so that Grok Code Fast could sit inside existing toolchains instead of asking developers to switch editors entirely.
The Benchmark Gap That Foreshadowed a Bigger Shift
xAI’s own internal evaluation harness put Grok Code Fast 1 at 70.8% on SWE-bench Verified, the widely used benchmark that tests whether a model can resolve real GitHub issues end to end. But independent evaluator Vals.ai, running the same benchmark, measured 57.6% — a 13.2 percentage-point gap between the vendor’s self-reported number and a third party’s reproduction. That discrepancy became something of a case study in how much evaluation harnesses, prompt scaffolding, and tool access can swing a coding model’s apparent competence, and it is a dynamic that now shadows almost every benchmark claim across the AI coding tools category, not just xAI’s.
The gap mattered for a simple reason: in a market crowded with coding assistants from Anthropic, OpenAI, Google, and a growing list of independent IDE makers, a self-reported benchmark number is only as useful as the trust developers place in it. When an outside lab’s result comes in more than 13 points lower, it reinforces a lesson the market had already been learning — that model marketing built around parameter counts or headline scores tells a smaller part of the story than how a tool performs on a team’s actual repository.
xAI Retires Grok Code Fast, Eight Models at Once
Less than nine months after launch, xAI pulled the plug. The company retired the grok-code-fast-1 API model on May 15, 2026, at 12:00 p.m. Pacific Time, as part of a broader cleanup that removed eight Grok models in total, including Grok 3, Grok 4-0709, several Grok Fast variants, and Grok Imagine Image Pro. A changelog dated May 8, 2026 reportedly told developers that Grok Code Fast 1 would disappear from Copilot Chat, inline edits, ask mode, agent mode, and code completions, pointing users toward GPT-5 mini and Claude Haiku 4.5 as stand-ins during the transition.
Rapid retirement cycles are not unusual in a market where model generations now turn over every few months, but the speed here — barely three quarters between launch and sunset — underscores how little patience xAI has shown for keeping an underperforming product on the shelf once a faster-moving successor exists.
Grok Build Steps In as the New Agentic Coding Model
The grok-code-fast-1 identifier was reportedly rerouted to grok-build-0.1, a purpose-built agentic coding model tied to xAI’s Grok Build CLI. The new model keeps the 256K context window of its predecessor but shifts the pricing structure: $1 per million input tokens, $0.20 per million cached-input tokens, and $2 per million output tokens. That is a higher headline rate than Grok Code Fast 1’s launch pricing, but the addition of a separate, cheaper cached-input tier signals xAI is optimizing for the repeat, multi-turn querying pattern that agentic coding workflows actually generate — where the same codebase context gets reread dozens of times across a single task.
The naming itself is telling. “Grok Build” frames the product around an outcome — shipping software — rather than around a benchmark score or a parameter count, which fits neatly with the “possibilities, not parameters” language now circulating about Musk’s broader approach.
Grok Bot Becomes a Router, Not Just a Bigger Model
The clearest signal yet came in October 2026, when Musk announced that Grok Bot — xAI’s consumer-facing agent service — would stop relying solely on Grok models and instead act as a router across multiple AI systems. According to Musk, simple requests would be handled by a fast Grok variant, while more demanding tasks would be routed to whichever backend model suits the job best, including Claude Opus 5.5, Midjourney, and Suno.
Musk described the move directly: “Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs,” he wrote on X, in an announcement describing Grok Bot’s shift to a router architecture. He followed up with the underlying principle driving the change: “Operating principle is to give Grok Bot users the best possible combination of speed & intelligence,” Musk said, explaining that straightforward requests would stay on a fast Grok variant while harder tasks get routed elsewhere.
That is a notable concession for a company that has spent much of the last two years touting Grok’s own scale. Rather than insisting every task run through the largest in-house model available, xAI is explicitly willing to hand work to a competitor’s model — Anthropic’s Claude Opus 5.5 — when it is the better tool for the job. It is hard to find a cleaner real-world example of “possibilities over parameters” than a company publicly routing traffic away from its own flagship when a rival’s system performs the task better.
Grok 4.7’s Parameter Jump Still Happened — Just Not as the Headline
None of this means xAI has abandoned scale entirely. Grok 4.7 launched in mid-September 2026 at roughly 2.1 trillion parameters, a 40% jump from Grok 4.6’s 1.5 trillion, with Musk claiming the model incorporates decades of SpaceX rocket, satellite, and manufacturing data. The difference now is framing: the parameter count is presented as backstory and provenance — what the model learned from — rather than as the headline metric buyers are told to care about. The actual purchasing decision facing developers today is pricing and context length, not trillions of parameters.
Cursor’s own pricing documentation, updated October 16, 2026, lists Grok 4.7 at $2 per million input tokens and $6 per million output tokens, jointly trained by Cursor and xAI’s infrastructure partners. Grok 4.7 Fast — the same model served on faster infrastructure — runs at double that rate, $4 input and $12 output per million tokens, with long-context requests above 256,000 tokens billed at a 2x multiplier up to a 500,000-token ceiling. Cursor’s documentation specifically notes that Grok 4.7 Fast is available only through Cursor and Grok Build, not through xAI’s public API.
| Model | Context Window | Input Price (per 1M tokens) | Output Price (per 1M tokens) | Status (as of Oct 19, 2026) |
|---|---|---|---|---|
| Grok Code Fast 1 | 256K | $0.20 | $1.50 | Retired May 15, 2026 |
| grok-build-0.1 | 256K | $1.00 ($0.20 cached) | $2.00 | Active |
| Grok 4.7 | Up to 500K | $2.00 | $6.00 | Active |
| Grok 4.7 Fast | Up to 500K (2x billing over 256K) | $4.00 | $12.00 | Active, Cursor/Grok Build only |
Macrohard and the Push Toward AI That Writes Software Itself
The strategic thesis behind all of this is Macrohard, Musk’s stated plan for an AI venture built around the idea that software agents — not exclusively human engineers — can design, write, test, and ship code. Musk has talked about this ambition in blunt terms. “We are doubling down on coding and making coding one of the highest priority efforts in the company,” he said during an xAI all-hands discussion, according to an archived recording circulated by outlets tracking xAI’s internal communications. He went further on timelines, predicting: “We do expect GROK code to be state of the art in two to three months,” a forecast that places the bar for xAI’s coding ambitions squarely in early 2027.
Musk’s most striking framing, though, goes past “better coding assistant” and into territory that sounds a great deal like the possibilities-over-parameters language now attached to his flying-car and supercar reveals. “The AI just creates the binary directly and the AI can create a much more efficient binary than can be done by any compiler,” Musk said, describing a future where AI-generated binaries could skip conventional source-code compilation entirely. If realized, that would not be an incremental improvement to developer tooling — it would be a different model of what “writing software” even means, which is precisely the kind of boundary-pushing claim that outlets like 36Kr are framing as the throughline across all of Musk’s current ventures.
Why the Compiler Claim Matters More Than It Sounds
Compilers have been treated as a settled, almost invisible layer of computing for decades — deterministic, auditable, and predictable in ways that make large engineering organizations comfortable shipping critical systems. Musk’s claim reframes that layer as just another target for AI disruption, alongside the model itself. Whether or not grok-build-0.1 or its successors can actually deliver more efficient binaries than a modern optimizing compiler is unverified and, per xAI’s own disclosures so far, remains a stated goal rather than a shipped capability. But the ambition itself is consistent with the “possibilities humanity has never achieved” framing that outlets are now attaching to Musk’s broader body of work.
A Timeline of xAI’s Coding Pivot
| Date | Event |
|---|---|
| August 28, 2025 | Grok Code Fast 1 launches as xAI’s first dedicated agentic coding model |
| May 8, 2026 | Changelog confirms Grok Code Fast 1 removal from Copilot Chat, inline edits, agent mode |
| May 15, 2026 | xAI retires grok-code-fast-1 along with 7 other Grok models |
| Mid-September 2026 | Grok 4.7 launches at ~2.1 trillion parameters, up 40% from Grok 4.6 |
| October 16, 2026 | Cursor updates pricing docs listing Grok 4.7 and Grok 4.7 Fast rates |
| October 2026 | Grok Bot becomes a multi-model router, pulling in Claude Opus 5.5, Midjourney, Suno |
How This Fits the Broader AI Coding Tools Market
xAI is not operating in isolation. The AI coding agent category has become one of the most contested corners of the software industry, with Anthropic’s Claude Code, OpenAI’s Codex, GitHub Copilot, Cursor, and a long list of smaller entrants all competing on a shifting mix of benchmark scores, pricing, context window size, and IDE integration depth. What makes xAI’s path notable is not that it is winning that race outright — the Vals.ai benchmark gap suggests it has real ground to cover — but that its public messaging has moved away from “we have the biggest model” and toward “we route you to whatever model gets the job done,” a position few of its rivals have been willing to state as plainly.
That is a meaningful departure from how the AI industry has generally marketed itself since the GPT-3 era, when parameter counts functioned almost like a spec sheet, similar to megapixels on a camera or horsepower on a car. A router model, by contrast, implicitly admits that no single model is the best choice for every task — an argument that is harder to put on a slide but potentially more honest about how production AI systems actually get used inside engineering teams today.
A Sample Grok Build API Call
For developers evaluating grok-build-0.1 against its predecessor, the request structure mirrors the OpenAI-compatible format xAI has used across its API since Grok Code Fast 1:
curl https://api.x.ai/v1/chat/completions \
-H "Authorization: Bearer $XAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "grok-build-0.1",
"messages": [
{"role": "user", "content": "Refactor this function for readability."}
],
"max_tokens": 4096
}'The cached-input pricing tier on grok-build-0.1 rewards exactly this kind of repeated, context-heavy querying pattern, which is standard in agentic coding sessions where the same repository context gets sent with every follow-up request.
Historical Context: From Parameter Races to Capability Races
The shift xAI is making did not happen in a vacuum. The broader large language model industry spent roughly 2020 through 2024 locked in a parameter arms race, with each new release touting bigger training runs as the primary evidence of progress. That framing started cracking once smaller, more efficient models began matching or beating larger predecessors on practical benchmarks, forcing vendors to compete on cost-per-task and real-world reliability instead of raw scale. xAI’s own Grok Code Fast to Grok Build transition, and now Grok Bot’s pivot to routing, is essentially that industry-wide trend playing out inside a single company’s product line over the span of about 14 months — from an August 2025 launch pitched on speed and integration, to a May 2026 retirement driven by benchmark credibility problems, to an October 2026 architecture that treats model choice itself as the product.
Market Impact for Developers and Enterprises
For engineering teams currently budgeting AI coding tool spend, the practical takeaway is that model selection is becoming a moving target even within a single vendor’s product line. Teams that built workflows or internal tooling around grok-code-fast-1’s $0.20/$1.50 pricing had roughly nine months before that model disappeared entirely, forcing a migration to grok-build-0.1 at a materially higher per-token rate — a five-fold jump on input pricing before accounting for the new cached-input discount. Any procurement process that assumes pricing or model availability will hold steady for a budget cycle is now making a riskier assumption than it would have been two years ago.
At the same time, Grok Bot’s router model points toward a future where the underlying model choice becomes less visible to end users and procurement teams alike — billed through a single subscription or API relationship even as requests get distributed across multiple vendors’ infrastructure behind the scenes. That has implications for data governance and compliance teams who need to know which company’s model actually processed a given request, a question that gets harder to answer cleanly once routing logic — rather than a fixed vendor relationship — decides where a prompt ends up.
What Musk and xAI Have Said, Verbatim
The clearest window into how seriously xAI is taking the coding push comes from Musk’s own statements, captured during internal discussions and public posts. On priority: “We are doubling down on coding and making coding one of the highest priority efforts in the company,” Musk said during an xAI all-hands. On timeline: “We do expect GROK code to be state of the art in two to three months,” he added in the same discussion. On the longer-term ambition to bypass traditional compilation entirely: “The AI just creates the binary directly and the AI can create a much more efficient binary than can be done by any compiler,” Musk said. And on the Grok Bot router shift specifically, Musk’s two posts — on using “the best back end model for any given task” and on the “operating principle” of balancing “speed & intelligence” — together lay out the clearest public statement yet of a company choosing task-fit over in-house model loyalty.
Predictions: Where xAI’s Coding Push Goes Next
- Expect grok-build-0.1 to face its own independent-benchmark test within weeks of wider developer adoption, given how much scrutiny the 70.8%-versus-57.6% gap drew for its predecessor.
- Musk’s two-to-three-month “state of the art” timeline for Grok Code points to an early-2027 checkpoint where xAI will face direct pressure to show SWE-bench or equivalent results that hold up under third-party review.
- Grok Bot’s router architecture is likely to expand to more backend partners beyond Claude Opus 5.5, Midjourney, and Suno, especially if usage data shows meaningful task-level performance gaps between in-house and third-party models.
- Pricing on grok-build-0.1 will likely see further tiering as agentic, cache-heavy workloads become the dominant usage pattern across the coding-assistant market, not just at xAI.
- If Musk’s compiler-bypass ambition for Macrohard produces even an early demo, it will intensify debate over code auditability and safety review processes that currently assume human-readable source as the artifact of record.
The Bigger Picture Behind the Headlines
Outlets framing Musk’s current moment around flying cars and supercars are not wrong that those products generate more public attention than a coding-model pricing change. But the coding-tools thread is arguably the more measurable test of the “not parameters, but possibilities” claim, because it comes with dates, dollar figures, and benchmark scores attached rather than a single product demo. A model that got retired after nine months, a benchmark gap an outside lab could not reproduce, a router that hands work to a competitor’s model when that is the better choice, and a stated goal of having AI write binaries directly rather than compile them — together, that is a company visibly trying to redefine what counts as progress in software, even as its parameter counts keep climbing in the background rather than out front.
Whether grok-build-0.1 or a future Grok Code release actually becomes “state of the art” on Musk’s self-imposed two-to-three-month clock is something only an independent benchmark run will settle. Until then, the more durable signal is the architecture itself: a consumer AI product willing to publicly route around its own flagship model, and a coding company treating a nine-month-old product’s retirement as routine business rather than a setback worth hiding.
Frequently Asked Questions
What happened to Grok Code Fast 1?
xAI retired grok-code-fast-1 on May 15, 2026, at 12:00 p.m. Pacific Time, as part of a cleanup that removed eight Grok models in total. The model’s API identifier was reportedly rerouted to grok-build-0.1, a successor agentic coding model.
How does grok-build-0.1 pricing compare to Grok Code Fast 1?
Grok Code Fast 1 launched at $0.20 per million input tokens and $1.50 per million output tokens. grok-build-0.1 is priced higher at $1.00 per million input tokens ($0.20 for cached input) and $2.00 per million output tokens, both with a 256K context window.
What is Grok Bot’s model router?
Announced by Elon Musk in October 2026, Grok Bot’s router automatically selects among multiple backend AI systems — including Claude Opus 5.5, Midjourney, and Suno — depending on the task, rather than relying solely on Grok models for every request.
How many parameters does Grok 4.7 have?
Grok 4.7 launched in mid-September 2026 at approximately 2.1 trillion parameters, a 40% increase over Grok 4.6’s reported 1.5 trillion parameters.
What is the gap between xAI’s self-reported benchmark and the independent score?
xAI’s internal harness scored Grok Code Fast 1 at 70.8% on SWE-bench Verified, while independent evaluator Vals.ai measured 57.6% on the same benchmark — a 13.2 percentage-point gap.
What is Macrohard?
Macrohard is Musk’s stated plan for an AI venture centered on the idea that software agents can design, write, test, and ship code with reduced reliance on human engineers, extending into ambitions like AI-generated binaries that could bypass traditional compilers.
Where can I access Grok 4.7 Fast?
According to Cursor’s pricing documentation updated October 16, 2026, Grok 4.7 Fast is available only through Cursor and Grok Build, not through xAI’s general public API, and is billed at double the standard Grok 4.7 token rate.
When did Musk say Grok Code would be state of the art?
During an xAI all-hands discussion, Musk said, “We do expect GROK code to be state of the art in two to three months,” placing the target window around early 2027 from the time of the statement.
