Grok 5 Launch Tracker: What AI Builders Should Watch
Grok 5 is still in training as of May 2026. Here's what's confirmed about xAI's next-gen model and what builders should track before switching.
Hello, I’m Dora. I’ve been refreshing the Grok 5 Polymarket contract more often than I’d like to admit. Not because I’m betting on it — because half my planning conversations this quarter have stalled on the same question: do we hold a slot for it or not.
So here’s the short version, before anything else. As of early June 2026, Grok 5 is still in training. It has not shipped. The prediction market currently prices a public release before June 30 at around 7%. xAI’s official line points to a Q2 window, but the original Q1 target already slipped, and “Q2” has quietly become “sometime.” That’s the whole situation in three sentences.
This isn’t a review — you can’t review a model that doesn’t exist yet. It’s a tracker. What’s confirmed, what’s reported, what’s a Musk tweet, and what any of it means for a production decision you have to make this month. I’ll tag every fact with where it came from, because the gap between “xAI said” and “someone on X claimed” is the entire game here.
Where Grok 5 stands today
xAI’s official position vs. community speculation
The only hard, sourced facts come from xAI’s January 28, 2026 Series E funding announcement, which confirmed Grok 5 was in active training on the Colossus 2 supercluster. Everything warmer than that — specific dates, benchmark leaks, AGI talk — is either Musk being Musk on X or community inference.
The timeline has moved more than once. The Q1 2026 release window Elon Musk originally targeted has passed; xAI’s official X account now points to Q2 2026 as the most likely window for a public beta. When the official channel and the founder’s posts disagree, I weight the official channel and treat the posts as mood, not roadmap.
Polymarket probabilities on a Q2 release
This is the number I’d actually anchor on, because it has money behind it rather than vibes. The Polymarket contract on a Grok 5 public release by June 30, 2026, currently sits around 7%. It’s been a volatile contract — it opened near 68 cents back in winter, collapsed into the low teens in early April on a string of repricings, then bounced through the 30s–60s on roadmap chatter before settling back down. The directional read across all that noise: the crowd treats a June ship as a coin-flip-or-worse, leaning worse.
Worth knowing what “Yes” even requires. The market resolves Yes only if Grok 5 is made publicly accessible — open beta or open rolling waitlist counts, but a closed beta or any private access does not. So a quiet internal milestone, even a real one, doesn’t move this contract. That’s a useful filter for your own planning, too.
V9-Medium training reports and the parameter claim
Here’s where the sourcing gets soft. Community reports — picked up in market context notes — describe a 1.5-trillion-parameter foundation model, internally tagged V9-Medium, finishing training in late May 2026, with fine-tuning and reinforcement learning underway and Musk suggesting a rollout within a couple of weeks. Two-to-three-week windows from Musk have a long history of not resolving in two to three weeks.
I’m flagging that whole paragraph as community reports/needs verification. Not dismissing it — V9-Medium may well be a real internal checkpoint. But “internal model finished a training run” and “Grok 5 is shipping to the public” are different events, and only the second one matters for your stack.
What’s claimed (and how confident we should be)
The spec sheet floating around is impressive. It’s also almost entirely unverified. Here’s how I’d sort it.
| Claim | Source type | Confidence |
|---|---|---|
| ~6T parameters, MoE architecture | Musk (Baron conf, late 2025) + reports | Reported — one founder mention |
| 1.5M-token context window | Community/aggregator reports | Reported — needs verification |
| Native multimodal input (text/image/audio/video) | Community reports | Reported — needs verification |
| In training on Colossus 2 | xAI official (Jan 2026) | Confirmed |
| June 2026 public release | Prediction market | ~7% probability |
The 6T-parameter and 1.5M-context claims
The 6 trillion parameter count is the only technical figure Musk confirmed directly, at the Baron Capital conference in late 2025, described as delivering higher intelligence density per gigabyte than simple scaling would suggest. One founder mention at an investor event is not a model card. It would use a Mixture-of-Experts architecture, meaning only a subset of parameters activate per query — so 6T total doesn’t mean 6T worth of compute per response. The 1.5M context figure shows up consistently across aggregator write-ups but traces to no primary xAI doc I can find. Treat both as reported.
The thing nobody phrases carefully enough
Multimodal input is not generation. Even if every leaked spec holds, Grok 5 is described as a model that receives text, images, audio, and video and outputs text. It is not an image or video generation model. If your pipeline needs to produce visuals, Grok 5 — whenever it lands — does not slot into that step. I keep seeing this conflated, and it leads teams to wait on a launch that was never going to touch their actual workload.
What xAI is shipping today instead
While the Grok 5 contract churns, xAI has been shipping the 4.x line steadily — and that’s what’s actually available to build on.
Grok 4.x as the current production option
The current flagship is Grok 4.3, released April 30, 2026, with a 1M token context, native video input, and native PDF/PPTX/XLSX file generation. That video-input capability is the part most relevant to builders: you can feed video to 4.3 today, without waiting. For most production needs that people are holding for Grok 5, 4.3 already clears the bar.
The pattern underneath matters too. xAI launched Grok 4.20 Beta on February 17, iterated to Beta 2 on March 3, with further betas in development. A team iterating this aggressively on the current line usually isn’t days away from replacing it — heavy 4.x investment is itself a soft signal that 5 isn’t imminent.
How to plan production work while Grok 5 is in training
My actual advice, since this is the part you can act on: don’t pause decisions for Grok 5. A 7% probability of a public release in the next several weeks is not a reason to freeze a roadmap.
When Grok 4.x is enough — which is most of the time — ship on it now. It’s available, documented, and stable. When Grok 5 is genuinely worth testing — a reasoning-heavy or long-context workload where a generational jump could change your architecture — the right move is to test it the day it’s publicly released, against your own held-out eval, not against launch-day benchmark headlines. Per Google’s own guidance on helpful, people-first content, the signal that ages well is real evaluation, not hype-chasing, and the same discipline applies to model selection.
What builders should track for confirmation
Signals worth watching, in rough order of how much they actually mean:
- Official xAI announcement + API documentation. Until xAI publishes it, it isn’t shipped. As Reuters’ ongoing xAI coverage and other primary reporting will confirm a real launch, treat consensus reporting plus xAI’s own channel as the bar — not a single tweet.
- Stable model IDs vs. preview suffixes. A -preview or -beta string in the API is a different commitment than a stable ID. Don’t build production routing against a preview alias.
- Published pricing and rate limits. This is the most underrated production signal. A model isn’t really shippable for you until you can see what it costs and how hard you can hit it. Pricing publication tends to lag the announcement, sometimes by weeks.
FAQ
Has Grok 5 been released?
No. As of early June 2026, Grok 5 remains in training. xAI confirmed active training in its January 2026 Series E update, and no public release has been announced since.
What’s the expected release window for Grok 5?
xAI’s official channel points to Q2 2026, but the original Q1 target already slipped. The Polymarket contract on a public release by June 30 currently sits around 7% — closer to “probably not this window” than a confident yes.
Does Grok 5 generate images or video?
Based on reported specs, no. Grok 5 is described as accepting multimodal input (text, image, audio, video) and producing text output. It is not an image or video generation model. If you need to produce visuals, this isn’t the tool — and that’s true even after it ships.
How does Grok 5 compare to GPT-5.5 or Claude Opus 4.7?
You can’t meaningfully compare an unreleased model to shipped ones. GPT-5.5 was released on April 23, 2026, and reset the bar Grok 5 would need to clear. Any Grok 5 vs. GPT-5.5 benchmark circulating now is based on leaks, not a public model. Wait for a real release and your own eval.
Should I integrate Grok 4 today or wait for Grok 5?
Integrate Grok 4.3 today if it meets your needs — it’s the current flagship with 1M context and native video input. Don’t stall a production decision on a model with a ~7% chance of shipping this window. Keep a fallback path so you can test Grok 5 on release day without re-architecting.
Conclusion
Strip away the AGI talk and the parameter-count theater, and Grok 5 is a model that’s still training, with a roughly one-in-four shot at a public release this window, that will output text and not pixels when it does arrive. That’s not a knock on it. It’s just where things actually stand.
The builder move is unglamorous: ship on Grok 4.3 now, keep your eval portable, and test Grok 5 the day it’s genuinely public — against your own data, not a launch thread. The teams that win the next launch cycle aren’t the ones who guessed the date right. They’re the ones who didn’t need to.
I’ll update this when xAI publishes something with an API doc attached. Until then — to be verified.
Previous posts: