The name Grok 4.6 entered the record as a three-word reply. Asked on 18 July 2026 whether the 2-trillion-parameter model he had just described was Grok 4.6, Elon Musk answered: "Yeah, Grok 4.6." Six days later he added a schedule. Searching SpaceXAI's own channels on 27 July 2026 turns up no third statement, and the figures attached to the model's name elsewhere were published about Grok 4.5.
What Elon Musk actually said about Grok 4.6
Two posts, six days apart, neither of them a press release.
On 18 July, replying to @minchoi, Musk wrote: "Our 2T model, which is better than our 1.5T in every way, will finish initial training next week. It might be able to exceed Kimi, but with speed and token efficiency close to our 1.5T (aka Grok 4.5)." He had not named the model. When Andrew Curran asked whether this was Grok 4.6, Musk confirmed it in the reply above.
Two details get lost in retelling. Musk said "Kimi," not "Kimi K3"; the identification with Moonshot's current flagship is a reasonable reading, not a quote. And the efficiency comparison is anchored to Grok 4.5, not to Kimi.
On 24 July came the timeline: "Grok 4.6 in 2 weeks and Grok 4.7 in 4 weeks." Two weeks from that date lands near 7 August. That post is itself a reply, and the parent matters: Musk had just posted a FrontierCode 1.1 score-versus-cost chart claiming "Grok 4.5 and Opus 5 are alone on Pareto frontier." The schedule was an answer inside a thread about cost efficiency.
That yields four confirmed items:
- Scale: approximately 2 trillion total parameters, up from Grok 4.5's 1.5T V9 foundation
- Capability target: possibly exceeding Kimi, whose K3 release Moonshot specifies at 2.8T total parameters
- Constraint: speed and token efficiency close to Grok 4.5
- Timing: about two weeks from 24 July, so early August
Architecture is undisclosed. Whether 2T is dense or mixture-of-experts, how many parameters activate per token, and whether V9 is extended or replaced are all unknown outside SpaceXAI. A throughput claim that survives a 33% parameter increase only means something once those details land.
Grok 4.6 is not in the SpaceXAI API yet
One naming note first, because it trips up searches. The company that builds Grok was founded as xAI in 2023 and now brands itself SpaceXAI after being folded under SpaceX in 2026. The documentation title reads "SpaceXAI Docs," the Grok 4.5 announcement calls it "SpaceXAI's smartest model," and Artificial Analysis lists the model as coming from SpaceXAI. The domains did not move: x.ai, docs.x.ai, and api.x.ai all still resolve, and model IDs are unchanged.
The SpaceXAI model documentation tags Grok 4.5 as Latest in the sidebar and New in the body. There is no grok-4.6 entry.
Six text models are documented as of 27 July 2026:
| Model ID | Context | Input / 1M | Output / 1M |
|---|---|---|---|
grok-4.5 | 500K | $2.00 | $6.00 |
grok-4.3 | 1M | $1.25 | $2.50 |
grok-4.20-0309-reasoning | 1M | $1.25 | $2.50 |
grok-4.20-0309-non-reasoning | 1M | $1.25 | $2.50 |
grok-4.20-multi-agent-0309 | 1M | $1.25 | $2.50 |
grok-build-0.1 | 256K | $1.00 | $2.00 |
Those are under-200K-context rates. Requests crossing 200K tokens bill at double for the entire request, so a 210K-token Grok 4.5 call costs $4.00 per million input and $12.00 per million output, not a blend of the two tiers.
No public SpaceXAI documentation or public model ID verifies Grok 4.6 access anywhere today. Treat any third-party offer of it as unverified unless the seller shows the exact SpaceXAI model ID and a dated API response. For production routing, that settles it: nothing changes until the model ID exists, and Grok 4.5 stays the current option.
Every performance number attached to Grok 4.6 belongs to Grok 4.5
Five figures recur across search results for the new model. Each was published about its predecessor:
| Figure | Actually describes |
|---|---|
| 80 transactions per second | Grok 4.5, SpaceXAI's launch claim |
| 29.0% on SWE Marathon, ahead of Claude Opus 4.8's 26.0% | Grok 4.5, SpaceXAI's launch claim |
| $2 input / $6 output per million tokens | Grok 4.5, current API price |
| 500K context window | Grok 4.5 |
| 1.5 trillion parameters, V9 foundation | Grok 4.5 |
For Grok 4.6, SpaceXAI has published no price, no context window, no benchmark score, no model ID, and no statement on vision, audio, or open weights.
The Grok 4.5 baseline, measured independently
Grok 4.6 will be judged against Grok 4.5, which makes the predecessor's measured numbers the only actionable data available before launch.
80 transactions per second is not 80 tokens per second
These are separate metrics that get quoted interchangeably. A transaction is a completed request. Output tokens per second measures generation rate within a single response. Neither converts to the other without knowing average response length.
Artificial Analysis measured Grok 4.5 at high reasoning effort at 58.3 output tokens per second, 102nd of 190 models, against a 73.0 t/s median for models in the same price band. Time to first token came in at 8.31 seconds versus a 2.80-second median. These are snapshot values read on 27 July 2026; the leaderboard re-tests continuously and rankings shift as models are added.
Intelligence is where Grok 4.5 earns its position: 54 on the same index, 13th of 190, against a median of 32 for comparable models. The profile is a smart model that starts late and generates slowly, priced below the $10.00 median for output.
Two release dates are also in circulation for Grok 4.5 itself. Artificial Analysis lists 8 July 2026; SpaceXAI's announcement page and contemporaneous coverage carry 16 July.
The cost floor Grok 4.6 has to clear
Cached input on Grok 4.5 runs $0.30 per million tokens, an 85% discount off the $2.00 input rate, which is the lever that matters for agent loops replaying long system prompts. Artificial Analysis reports a $1.35 per million blended rate under its own methodology, which assumes a 7:2:1 mix of cache hits, uncached input, and output.
Whether SpaceXAI holds these rates at 2T is one input into whether the Pareto-frontier claim in that parent post survives its own successor. Benchmark score, latency, and context all move that line too, and none of them are known for 4.6 yet.
Why "2T at Grok 4.5 speed" is a harder promise than it sounds
Read literally, the constraint is modest. Grok 4.5 already generates below the median rate for its price tier and takes nearly three times the median to produce a first token. Holding that line at 2T means preserving a slow-but-cheap profile, not defending a speed record.
Read against the competitive target, it gets harder, though the headline parameter gap is less decisive than it looks. Moonshot specifies Kimi K3 at 2.8T total parameters in a sparse mixture-of-experts design that activates 16 of 896 experts per token, so its inference-time cost tracks the active slice rather than the full 2.8T. SpaceXAI has not said whether Grok 4.6's 2T is dense or sparse, which is exactly the detail that determines whether these two numbers are comparable at all. K3 also lists a 1M-token context and open weights at $3 input and $15 output per million tokens, against Grok 4.5's 500K, closed weights, and $2/$6.
Neither side of that has been measured. Grok 4.6 has not been benchmarked by SpaceXAI or by anyone else.
How to tell the day Grok 4.6 actually ships
Check the documentation and the live endpoint, in that order:
1. A grok-4.6 entry appears at docs.x.ai/developers/models with its own context window and price row. 2. The API model list returns the new ID. One request settles it:
curl https://api.x.ai/v1/models \
-H "Authorization: Bearer $XAI_API_KEY" | grep grok-4.6A returned ID confirms the model exists; it does not confirm your account is entitled to it, so check rate limits and pricing on your own key before routing anything.
3. SpaceXAI publishes a model card with benchmark numbers. Treat vendor benchmarks as the floor and wait for third-party evaluation before rewriting routing rules.
Distribution beyond the API has not been announced. SpaceXAI's published plans cover the free tier, SuperGrok, Business, and Enterprise, and Grok 4.6 reaching those alongside the API is an expectation drawn from how 4.5 rolled out, not a commitment from SpaceXAI.
Production routing decisions can wait: Grok 4.5 remains SpaceXAI's flagship at a known price with a known context window, and Musk's own chart places it on the cost-versus-score frontier next to Claude Opus 5. Planning decisions are different. If you are budgeting for Q4 inference or negotiating a contract, an early-August 2T release at unknown pricing is a variable worth holding open. Two weeks is also a schedule Musk set for himself.
FAQ
When is Grok 4.6 coming out?
No official date exists. Musk said on 24 July 2026 that Grok 4.6 would arrive in about two weeks, putting it near 7 August, with Grok 4.7 about two weeks after that. SpaceXAI has not published a launch date on any of its own channels.
How many parameters does Grok 4.6 have?
Roughly 2 trillion total, per Musk on 18 July 2026. Grok 4.5 sits at 1.5T on the V9 foundation. Whether Grok 4.6 is dense or mixture-of-experts, and how many parameters are active per token, has not been disclosed.
How much will the Grok 4.6 API cost?
Unannounced. Grok 4.5 costs $2.00 per million input tokens and $6.00 per million output under 200K context, doubling above that threshold, with cached input at $0.30. Those rates are the reasonable planning baseline until SpaceXAI publishes actual figures.
Is Grok 4.6 better than Kimi K3?
Untested by anyone. Musk said the 2T model "might be able to exceed Kimi" while staying close to Grok 4.5 on speed and token efficiency, and he did not name K3 specifically. Moonshot specifies K3 at 2.8T total parameters with 16 of 896 experts active per token; SpaceXAI has not disclosed Grok 4.6's architecture, so the two parameter counts are not yet comparable.
Will Grok 4.6 come to Android and iOS?
Not announced. The Grok apps carry SpaceXAI's current models, so app access is the expected path once the model ships, without a stated date. Regional timing is a separate variable: Grok 4.5's European Union rollout was reported as still pending on 13 July 2026, and SpaceXAI has not said whether 4.6 ships globally on day one.
Is Grok made by xAI or SpaceXAI?
Both names refer to the same company. It was founded as xAI in 2023 and now brands as SpaceXAI following its move under SpaceX in 2026. Grok 4.6 is a SpaceXAI model. The x.ai domains and the grok-* model IDs were not renamed.
What is the latest Grok model right now?
Grok 4.5, released in July 2026 and tagged Latest in SpaceXAI's documentation as of 27 July. It offers a 500K context window, configurable reasoning, and text plus image input.
Related reading
*Pricing and model availability verified against SpaceXAI documentation on 27 July 2026.*