Grok 4.6 and 4.7 Timeline Puts Efficiency on Trial

Elon Musk said SpaceXAI will ship Grok 4.6 around August 7 as a 1.5 trillion-parameter model with significantly improved supervised fine-tuning and reinforcement learning. A 2.1T Grok 4.7 follows a few weeks later that he called better in every way except slightly slower to serve, with even better token efficiency.

The dual timeline lands days after Grok 4.5 topped an independent cybersecurity price-performance benchmark and only weeks after Moonshot opened the weights on its 2.8T Kimi K3. August will show whether continuous shipping and efficiency can move the frontier economics.

Musk Spells Out the Next Two Drops

Replying on X to Vercel CEO Guillermo Rauch, Musk laid out the schedule in plain terms.

Interesting. Grok 4.6 releases around August 7. This will be the 1.5T model with significantly improved SFT & RL. Grok 4.7 will be the 2.1T model released a few weeks later. This will be better than 4.6 in every way, except slightly slower to serve, albeit with even better token efficiency.

The post quoting Rauch’s cybersecurity note has drawn millions of views. Earlier Musk comments had pointed to a roughly two-week then four-week cadence from late July, so the August 7 window matches the tighter schedule. SpaceXAI has been shipping major Grok updates on a near-monthly rhythm.

The Grok 4.6 releases around August 7 claim gives developers a concrete near-term target. Parameter counts mark a clear step up from the current flagship while the SFT and RL upgrades target the training stages that often decide real-world agent and coding reliability.

That two-step plan also splits risk. The first drop holds the 1.5T class and focuses on training quality. The second adds raw scale a few weeks later. Teams that need a quick lift can plan around early August; teams that can wait get the larger model before summer ends.

The near-monthly rhythm itself is now part of the story. Holding that pace while climbing from the current flagship through 1.5T and then 2.1T would keep pressure on labs that ship on slower cycles.

Grok 4.5 Already Won on Cyber Cost

Rauch’s assessment of Grok 4.5 supplied the prompt for Musk’s reply. In Vercel’s latest DeepsecBench runs on real open-source vulnerabilities, Grok 4.5 emerged as the best cybersecurity AI model on price-performance.

• 10x cheaper than GPT-5.6 Sol per run
• 5.7x cheaper than Opus 5
• 2.2x cheaper than Kimi K3 while matching Kimi-like performance
• Sol still led raw accuracy, ahead of Opus 5

Reported Deepsec numbers put Sol near a 35.58 score at roughly $56 per run. Grok 4.5 landed in the 15.58-16.54 range at $5.60-$11 per run. For high-volume security scanning that gap decides whether the tool is usable at scale. One independent note flagged a higher general hallucination rate for Grok 4.5 versus prior versions, so teams still need to validate on their own codebases before production use.

Grok 4.5 itself launched in mid-July as SpaceXAI’s strongest coding and agentic model to date. Official materials list it priced at $2 per million input tokens and $6 per million output tokens, with claimed roughly 2x token efficiency versus comparable leading models and serve speeds around 80 tokens per second. Training included heavy Cursor collaboration and data spanning science, engineering and math. The model’s knowledge cutoff of February 1, 2026 is listed in the developer docs.

Those list prices and the Deepsec run costs tell the same story from two angles. Per-token rates set the floor; per-run totals show what a full vulnerability pass actually burns. The 10x and 5.7x gaps against Sol and Opus 5 are what make continuous scanning budgets suddenly realistic for more teams.

Raw accuracy still favors Sol. Price-performance favors Grok 4.5. That split is exactly why the August upgrades matter: if SFT and RL lifts close some of the accuracy gap without breaking the cost line, the middle of the market moves.

How the Models Stack on Paper

Public figures and the new roadmap put the current and next Groks in a clear line against the names Rauch cited.

Exact pricing and final scores for 4.6 and 4.7 remain unknown until release. The pattern SpaceXAI has set favors keeping absolute cost low while lifting capability through training improvements and scale.

Read down the parameter column and the path is linear: hold the 1.5T class, improve training, then step to 2.1T. Read across the notes column and the bet is different. Efficiency and serve behavior matter as much as size. Kimi K3 still sets the open-weight ceiling at 2.8T MoE, so the closed stack has to stay cheap enough that self-hosting is optional rather than required.

What the SFT, RL and Efficiency Upgrades Target

Supervised fine-tuning and reinforcement learning are the stages that turn raw pre-training into reliable multi-step behavior. Musk’s note that both improve “significantly” on the 1.5T 4.6 model points at better grading on long agent rollouts and engineering tasks, the same areas Grok 4.5 already stressed.

Grok 4.7’s 2.1T size is a straightforward capacity increase. The explicit trade-off is slightly slower serving offset by higher token efficiency. In practice that can mean fewer tokens burned per successful coding or security pass, which compounds on high-volume workloads even if wall-clock latency per request rises a little.

Crowd reaction on X quickly zeroed in on the cadence itself. Users noted a new frontier-class model roughly every month and argued that holding speed and efficiency near the 4.5 level while climbing parameters would make the stack the default for cost-sensitive coding and scanning. That expectation now sits on the August calendar.

Broken into parts, the stated upgrades land on three levers teams already measure:

• SFT and RL quality on the 1.5T 4.6 drop, aimed at agent rollouts and engineering reliability
• Parameter scale on the 2.1T 4.7 drop, a direct capacity step
• Token efficiency on 4.7, meant to offset slightly slower serve times on volume jobs

Grok 4.5 already claimed roughly 2x token efficiency and serve speeds around 80 tokens per second at $2/$6 per million tokens. Holding near that cost envelope while lifting SFT, RL, and later scale is the bar the next two releases have to clear. Miss the efficiency piece and the larger model becomes a niche tool. Hit it and the cost-per-solved-task curve keeps bending down.

Open Chinese Weights Meet Closed U.S. Cadence

Moonshot’s Kimi K3 arrived via API in mid-July and open weights landed around July 26-27 as a 2.8T mixture-of-experts model. Independent coverage described it as near-frontier on several suites while being easier to run than pure dense peers. That release raised the floor for anyone with GPU racks and simultaneously raised pressure on closed U.S. labs to justify premium pricing.

OpenAI and Anthropic have separately pushed for clearer federal review frameworks before advanced model releases. The Trump administration has been developing voluntary national-security vetting windows. xAI’s rapid public drops sit outside the slower, more cautious posture some peers have adopted. NVIDIA and other firms have also warned against broad open-weight restrictions, a stance that drew support from Musk and Sam Altman.

The practical collision is simple. High-volume users can already route cyber and coding jobs to Grok 4.5 or self-host Kimi-class weights. A stronger 4.6 and a larger, more efficient 4.7 give them two more closed options that still aim for low cost per useful token.

Open weights change the negotiation. A lab that charges a steep premium must now show accuracy or safety gains that self-hosted 2.8T MoE weights cannot match. A lab that ships fast and keeps prices near the Grok 4.5 line can argue that managed endpoints still win on convenience even when the weights are free elsewhere. August tests both arguments at once.

Who Feels the August Pressure First

Security teams running continuous vulnerability scans gain the most immediate leverage. A 5-10x cost drop for usable detection changes budget math overnight. Coding agents and IDE integrations are next; the earlier Grok 4.5 public release and coding focus already positioned the model inside Cursor and Grok Build. Faster iteration on SFT/RL should tighten that loop.

Closed frontier labs face a margin and cadence test. If 4.6 and 4.7 deliver near-4.5 speed and cost with higher capability, volume customers have less reason to stay on the most expensive endpoints for every task. Self-hosters and open-weight hosts gain another comparison point against Kimi K3.

Regulators and enterprise risk teams watch the hallucination and safety numbers. Grok 4.5’s reported rise in general hallucination rate is a live caution for any security deployment. 4.6’s improved RL will be judged on whether it tightens precision without losing the cost edge.

1. Mid-July 2026, Grok 4.5 launches publicly at $2/$6 per million tokens with heavy coding and agent training.
2. July 26-27 2026, Moonshot releases Kimi K3 open weights (2.8T MoE).
3. July 27-28 2026, Rauch posts DeepsecBench results naming Grok 4.5 price-performance leader; Musk replies with 4.6/4.7 timeline.
4. Around August 7 2026, Planned Grok 4.6 (1.5T, improved SFT & RL).
5. Late August / early September 2026, Planned Grok 4.7 (2.1T, broader gains, efficiency focus).

That sequence compresses what used to be multi-quarter jumps into a single summer. Developers will run the new models on the same Deepsec-style and SWE-style suites within days of each drop. The scores and the dollar-per-solved-task numbers will travel faster than any marketing claim.

Why Two Models Land Weeks Apart

The schedule is not a single leap. It is a 1.5T training-quality drop followed quickly by a 2.1T scale-and-efficiency drop. That order lets SpaceXAI ship the SFT and RL gains first, while the larger model is still finishing, and still keep both inside the same summer window.

For buyers the split creates a simple decision tree.

• Need better agents and coding reliability as soon as early August? Plan on 4.6 at the same 1.5T class.
• Need maximum capability and can trade a little serve latency for token efficiency? Wait for 4.7 a few weeks later.
• Already happy with Grok 4.5 price-performance on Deepsec? Keep routing there until independent scores land on the new drops.

The same split also limits downside. If 4.6’s training lifts underwhelm, 4.7 still offers a second chance on scale and efficiency. If 4.7’s serve speed disappoints, 4.6 remains the faster path at the prior size. Either way the near-monthly cadence keeps a fresh comparison point in front of Sol, Opus 5, and Kimi K3.

August Becomes the Efficiency Stress Test

If Grok 4.6 lands on schedule with clear gains on agent and cyber tasks at 4.5-like cost, the conversation shifts from “can xAI catch the leaders” to “how long can premium pricing hold for volume work.” Grok 4.7 then has to prove that a 2.1T model can stay practical to serve. Slight latency for better tokens-per-result is a trade many production stacks will accept.

Failure modes are equally plain. Slippage past early August, weak SFT/RL lifts, or efficiency numbers that do not materialize would let Sol and the strongest Claude variants keep their accuracy premium without immediate pressure. Kimi-class open weights would remain the default low-cost alternative for anyone willing to run their own iron.

The next six weeks therefore carry a clean experiment. Two larger Groks, one right after the other, measured against a fresh cybersecurity price-performance win and a newly open Chinese heavyweight. The market will price the outcome in API bills and default model choices long before the next policy framework is finished.

Volume buyers will not wait for white papers. They will compare dollar-per-solved-task on the same suites that already favored Grok 4.5 on price-performance, then route traffic accordingly. That is the stress test August actually runs.