Grok 4.5 Leak: What the Canary Traces, Cursor Training Data, and 1.5T V9 Foundation Actually Tell Us

On July 6, 2026, a string reading "Unlock the full power of Chat with Grok 4.5" surfaced in the Grok web UI. Not a launch. Not a system card. Not a keynote. A canary trace, spotted by TestingCatalog, that pushed a three-week trickle of xAI rumors into something concrete.

TLDR Grok 4.5 is xAI's next release, built on a 1.5T-parameter V9 foundation model with Cursor data mixed in during supplemental training. It has been in private beta at SpaceX and Tesla for roughly a week and, per Elon Musk, is "performing close to or beyond Opus" in internal evaluations. Traces have now appeared in the Grok web UI. No official benchmarks, pricing, context window, or release date have been published. A larger 2T V-series run is reportedly still cooking, aimed at an August release.

Key Takeaways

• Grok 4.5 is built on the V9 1.5T foundation model — approximately 3× the size of xAI's existing v8-small (~0.5T) per a summary by Rohan Paul.
• Cursor SFT + RL data was added during supplemental training, not pre-training — a distinction that matters for how much the model's agentic-coding gains will generalize.
• Private beta at SpaceX and Tesla since roughly June 28, 2026, per remarks from Musk relayed by TestingCatalog.
• Musk described early performance as "close to or beyond Opus" — an internal impression, not a published benchmark.
• A separate 2T-parameter V-series run reportedly started several weeks earlier and is expected to finish in late July for an August release.
• Zero benchmark numbers, pricing, context window sizes, or API details have been officially disclosed in the current signal set.

What Was Actually Seen

The picture assembled from the last three weeks of xAI signals is unusually specific for a leak cycle, and it's worth listing the concrete data points before the speculation.

• Canary UI trace, July 6, 2026. A "Grok 4.5" reference appeared inside the Grok web UI — surfaced text: "Unlock the full power of Chat with Grok 4.5" — per TestingCatalog. The tweet includes a screenshot of the string in the interface.
• Version scrub in menus, June 28. Mark Kretschmann noted that the version number had been removed from Grok/Cursor menus, writing that "this always happens shortly before a release from @xai". That's a pattern-match, not a confirmation.
• Foundation model: V9, 1.5T parameters. Elon Musk's public update, summarized by multiple observers, states Grok 4.5 is based on xAI's V9 1.5T foundation, with Cursor data added in supplemental training. See TestingCatalog's summary and Rohan Paul's follow-up.

Source: @testingcatalog

• Roughly 3× larger than v8-small. Rohan Paul's write-up puts the existing v8-small at ~0.5T parameters, making Grok 4.5's V9 foundation "approximately 3× larger."
• Private beta at SpaceX and Tesla. Multiple observers, including Ashutosh Shrivastava and TestingCatalog, reported Musk's quote: "Grok 4.5 is now in private beta at SpaceX and Tesla, performing close to or beyond Opus."
• Cursor team involved in v9 SFT and RL. Mark Kretschmann's summary of Musk's update attributes "excellent training data and major engineering work to v9 SFT and RL" to the Cursor team.
• A separate 2T run is in progress. Per the same summary, a new 2T-parameter run began several weeks earlier, is expected to finish in late July, and is targeted for an August release.

Source: @rohanpaul_ai

That is the full set of load-bearing facts. Everything below is context or interpretation.

Why the 1.5T V9 Foundation Matters

The most cited number in this cycle is 1.5T parameters. On its own, that number would just be a scale marker. The interesting part is what sits underneath it.

xAI appears to be running a versioned foundation-model track — V-series — with distinct 0.5T, 1.5T, and 2T runs progressing in parallel or in sequence. The current 1.5T run underlies Grok 4.5. The 2T run, still in training, is expected to power a subsequent release. That kind of laddered foundation strategy is what you'd expect if the lab is treating post-training and data mixture (SFT, RL, tool-use adaptation) as the primary lever, and pre-training scale as a periodic step function.

Haider's read of Musk's remarks puts it directly: xAI plans to "release new scratch-trained models every month this year," which "shows how much recent progress is coming from post-training." Treat the specific monthly cadence as aspirational, but the underlying claim — that post-training is where the delta is — matches what most frontier labs are signaling in mid-2026.

The Cursor Supplemental Training point is subtler than "the model was trained on Cursor data." Per Kretschmann's summary, Cursor data was added during supplemental training for the current 1.5T V9 run — meaning it was not blended into the base pre-training corpus, but layered on top through SFT and RL. That distinction has real implications:

• Coding-benchmark gains from supplemental data tend to be sharper on the tasks the data resembles (real IDE traces, tool-use sequences, edit patterns).
• Gains tend to generalize less to fundamentally out-of-distribution reasoning tasks.
• The base model's non-coding capabilities are less likely to be perturbed than they would be under pre-training-scale infusion.

For builders evaluating Grok 4.5 in the wild, that reads as: expect strong showing on agentic-coding evals that resemble Cursor's actual workflows, and be cautious about extrapolating those numbers to unrelated reasoning tasks.

The "Close to or Beyond Opus" Claim

Every leak cycle produces one quotable line. This one's is "performing close to or beyond Opus." It's worth unpacking.

The line originates in Musk's private-beta update, relayed via at least three accounts on the same day: Ashutosh Shrivastava, TestingCatalog, and 小互. It is an internal impression from SpaceX and Tesla evaluators, not a published benchmark. No SWE-bench score, MMLU number, GPQA figure, or agentic-coding eval accompanies it in this signal set.

What the community claim tells us: xAI's internal evaluators believe the model is competitive with Claude Opus on the tasks they care about. What it doesn't tell us: which Opus (there are multiple in circulation), on which benchmarks, at what latency, with what test-time compute budget, or on which subset of tasks the "close to" applies. Bindu Reddy's independent read ("hearing great things… fantastic to get a fast agentic model") adds directional support but no measurement.

Reasonable interim posture: the model is plausibly frontier-competitive on agentic coding, and probably not a full generational leap on general reasoning. Wait for third-party evals before adjusting production stacks.

Grok 4.5 vs Claude Opus: What the Signal Says

Claude Opus is the only named competitor Musk directly referenced. The comparison in the current signal set is thin, but here's what can be responsibly stated on each dimension.

• Parameter count. Grok 4.5 is on a 1.5T V9 foundation per Musk's public remarks. Claude Opus parameter count is unverified — no public number from either lab in this signal set.
• Coding orientation. Grok 4.5 layers Cursor SFT + RL data on top of the base model. Claude Opus's post-training recipe is not addressed in this signal set.
• Benchmarks published. Grok 4.5: none. Claude Opus: unverified — no public number from either lab appears in this signal set. The "close to or beyond Opus" line is a community impression, not a measured comparison.
• Availability. Grok 4.5: private beta at SpaceX and Tesla, per TestingCatalog. Opus: unverified in this signal set.
• Context window. Both unverified. No public number appears in this signal set for either side.

Do not read this table as an even head-to-head. Read it as a boundary condition: the only comparison that exists today is an internal quote, and that's not a benchmark.

What We Know vs. What We Don't

Separating signal from speculation is what makes this cycle usable for planning. The following bullets map 1:1 to the FAQ section of this article's schema.

What we know (from the bundle):

• Grok 4.5 is built on xAI's 1.5T V9 foundation model, with Cursor data added during supplemental training (per Musk, relayed by multiple observers).
• Grok 4.5 has entered a private beta at SpaceX and Tesla (TestingCatalog).
• Musk described Grok 4.5 as "performing close to or beyond Opus" in internal evaluations — an internal impression, not a published benchmark.
• Traces of Grok 4.5 have appeared in the Grok web UI on July 6, 2026, including the string "Unlock the full power of Chat with Grok 4.5" (TestingCatalog).
• Grok 4.5 is approximately 3× larger than the existing v8-small model (~0.5T parameters), per Rohan Paul's summary.
• A separate 2T-parameter V-series run reportedly started several weeks earlier and is expected to finish in late July for an August release (Mark Kretschmann's summary).

What we don't know (open questions):

• No official benchmarks. No SWE-bench, MMLU, GPQA, or agentic-coding eval numbers for Grok 4.5 appear in this signal set.
• No official release date. Kretschmann speculated "this week" as of July 6, 2026, but xAI has not confirmed a shipping date.
• No pricing, context window, rate limits, or API availability details.
• Cursor Composer 3 relationship is undefined. Kretschmann has referenced a "Grok 1.5T / Cursor Composer 3" model in the same breath, but the exact product mapping between the Grok 4.5 release and any Cursor product update is unclear.
• No licensing signal. Nothing in the bundle addresses open weights, commercial terms, or acceptable-use policy changes.
• Cursor data usage mechanics. Per Kretschmann's relay of Musk's remarks, Cursor data was added during supplemental training rather than pre-training — but the exact SFT/RL mixture and volume is not disclosed.

What Builders Should Do Today

Three concrete actions while the signal window is open.

Prepare an internal Cursor-shaped eval. If Grok 4.5's coding gains come from Cursor SFT and RL, they will be strongest on tasks that resemble Cursor-native workflows: multi-file edits, tool-use sequences, diff-oriented reasoning, agentic loops. Have a private eval ready that mirrors your actual codebase and tooling — not just SWE-bench. That way, on release day, you can measure the gap between "close to or beyond Opus" and your production reality within hours instead of weeks.

Do not migrate coding pipelines on the Musk quote alone. "Close to or beyond Opus" is an internal impression. Wait for at least one third-party eval (independent SWE-bench Verified run, a hands-on Reddit stress test, or an agentic benchmark from a lab not affiliated with either xAI or Anthropic) before adjusting production stacks. The distance between "internal impression" and "measured comparable" has historically been large.

Track the 2T run separately. The 2T V-series run, targeted for an August release per the Kretschmann summary, is a different animal from the 1.5T Grok 4.5. If you're planning capacity or vendor strategy for Q4, the 2T is the more relevant checkpoint. Keep a note that this second shoe is expected to drop and don't conflate the two.

The Week Ahead: What to Watch

Three concrete signals to monitor between now and whenever the release lands.

• Watch the model card. An official xAI model card — with an architecture summary, training-data disclosure, and at least a partial benchmark table — is the earliest artifact that will let independent evaluators start real comparisons. Any release without one is a yellow flag.
• Run your own coding eval before relying on the Opus comparison. The "close to or beyond Opus" claim is worth exactly zero for your codebase until you measure it on your codebase. Have the harness ready before day zero.
• Pin the 2T-run timeline. If Kretschmann's summary of a "late July finish, August release" holds, you'll get a second data point six to eight weeks after Grok 4.5. That second run is where you'll actually learn whether xAI's V-series strategy is compounding or plateauing. Watch for a fresh Musk post or another canary trace matching a new version string.

The next 72 hours will likely bring either an official announcement or a wider set of UI traces. Either way, the substance of Grok 4.5 will only be legible once real benchmarks — from anyone — start landing.

Building similar agentic-coding stacks with frontier LLMs? On kie.ai you can try Claude Opus 4.8, GPT-5.5, and Gemini 3 Pro.