AI Model Rumors and Release Roundups

Dated sourcing, reliability notes and clearly labelled expectations. A rumor score describes sourcing confidence, not a confirmed model specification.

Grok 5: Everything We Know

xAI · Expected Q1 2027 · Last updated 2026-10-24

Confidence: Rumored · Rumor score 68/100

Grok 5 is in training on Colossus 2 and is rumored to be a ~6-trillion-parameter mixture-of-experts, the largest announced model, with native multimodal input including video temporal reasoning and a rumored ~1.5M-token context. Earlier public-beta estimates for March–April 2026 slipped; the current expected window is Q1 2027.

Expected features

  • ~6T-parameter MoE (rumored, largest announced)
  • Native multimodal, incl. video temporal reasoning
  • ~1.5M-token context (rumored)
  • Trained on Colossus 2 (~550K GB200/GB300 cluster)

Architecture

Reports point to a ~6-trillion-parameter MoE trained on Colossus 2. Treat parameter counts as rumored until xAI publishes a model card.

Release Date

Public-beta expectations for Mar–Apr 2026 slipped. Trade press now points to Q1 2027. No date is confirmed.

Pricing

No pricing has been announced. Grok 4.7 held Grok 4.6 pricing at $2/$6; whether Grok 5 holds that line is unknown.

Reports and sourcing

Colossus 2 training run under way

xAI statements · High, on the record · 2026-10-24

xAI has said Grok 5 is training on Colossus 2. Scale and timing beyond that are not official.

~6T parameters, video temporal reasoning

Musk posts / trade press · Medium, directional, not a spec sheet · 2026-10-24

The 6T MoE figure and ~1.5M context circulate widely but remain unconfirmed by any xAI document.

Q1 2027 window

Trade press consensus · Medium · 2026-10-20

Multiple outlets moved expectations to Q1 2027 after the spring beta window passed.

Related reporting

Read the xAI family guide

Gemini 4 Pro (full Gemini 4 family): Everything We Know

Google · Expected No confirmed date; earlier late-Q4 speculation · Last updated 2026-11-01

Confidence: Rumored · Rumor score 74/100

Google announced Gemini 4 Carbon on 28 October with restricted access, following Argon's Stable GA. That is a confirmed model announcement, not confirmation of the separately rumored Pro name, feature list or December date. The earlier October Pro prediction has expired without a verified Pro launch; broader Carbon access has no announced date.

Expected features

  • A separately named Pro tier is unconfirmed
  • Pro reasoning and agent features remain speculation
  • Workspace integration is unconfirmed for a Pro release
  • Carbon is official; public access has no announced date

Features

Those themes were earlier expectations, not an announced Pro specification. Carbon's official announcement should not be relabeled as a Pro confirmation.

Release Date

Google has not announced a Pro release date. The old late-Q4 estimate was pattern-based speculation; the October anonymous prediction has now expired.

Pricing

Unknown for the Pro tier. Note the separate, confirmed changes on Google's price table: Gemini 4 Argon steps from $2/$10 to $4/$20 on 1 Dec 2026, and Gemini 3.8 Flash doubles on 1 Jan 2027.

Reports and sourcing

Annual cadence points to late Q4

Frontrumors analysis of release history · Medium-high, pattern, not a leak · 2026-10-20

Gemini 1/2/3 all landed late in their years; Argon's Sep announcement leaves room for a Pro tier before year-end.

Argon positioned as 'first Gemini 4 model'

Google announcement language · High, public wording · 2026-09-30

Google called Argon the first Gemini 4 model, implying more of the family follows.

Single-source insider claim of an October release

Anonymous X account relayed by NokiaPowerUser, 5 Sep 2026 · Low, one anonymous source, no corroboration · 2026-09-05

The post expected a public Pro release in October. October has passed without a verified Pro launch. Google announced Carbon on 28 October, which does not confirm this anonymous Pro prediction.

Related reporting

Read the Google family guide

Muse Wonder, GA: Everything We Know

Meta · Expected Q4 2026 – Q1 2027 · Last updated 2026-11-01

Confidence: Rumored · Rumor score 61/100

Muse Wonder entered preview on 12 October. Artificial Analysis now distinguishes Wonder and Wonder Full on Intelligence Index v5; the current dated measurements are in Benchmarks. GA timing remains unpublished and we could not verify first-party API pricing; Q4 2026–Q1 2027 is our estimate, not a Meta commitment.

Expected features

  • Separate Wonder and Wonder Full preview variants
  • Most capable Meta model to date
  • No first-party API price verified (OpenRouter lists $4/$20)
  • Preview access only as of 12 Oct 2026

What We Know

Wonder is real and in preview, that part is fact, not rumor. Only the GA date and pricing are rumored.

Release Date

Q4 2026–Q1 2027 is an estimate based on typical preview-to-GA gaps. Meta has published no date.

Pricing

We could not verify a first-party price. OpenRouter has listed muse-wonder at $4/$20 since 12 Oct; for reference, Muse Spark 1.3 lists at $1.25/$4.25 and Wonder should sit above it.

Reports and sourcing

Preview launch with frontier scores

Meta preview release, 12 Oct 2026 · Confirmed · 2026-10-12

Wonder previewed at ~52.4 Artificial Analysis index, level with GPT-6 Astra.

GA in Q4 2026 – Q1 2027

Frontrumors estimate · Low-medium, no official date · 2026-10-12

No Meta statement pins a GA window; treat any specific month you read elsewhere as speculation.

Two Wonder variants on the AA leaderboard

Artificial Analysis leaderboard · High, independent measurement · 2026-11-01

Intelligence Index v5 distinguishes Wonder (max) from Wonder Full (max). See Benchmarks for the current snapshot; older index values are not the same scale.

Related reporting

Read the Meta family guide

GPT-6.1 / next GPT point release: Everything We Know

OpenAI · Expected Dec 2026 – H1 2027 · Last updated 2026-11-01

Confidence: Rumored · Rumor score 57/100

GPT-6.1 Sol shipped on 29 September and OpenAI confirms availability in paid Work/Codex plans and the API. Terra followed on 26 October as a $1/$5 computer-use tier. A separate Astra-tier 6.1 refresh is still an unconfirmed rumor; historical October rollout reports are not evidence that Sol remains limited today.

Expected features

  • Astra-tier point refresh (rumored)
  • Wider GPT-6 availability across plans
  • 'GPT-6 Pro' tier inside ChatGPT Pro
  • Watch: GPT-5.6 Sol promo pricing is guaranteed only through at least 21 Nov 2026

Rollout Status

Confirmed: GPT-6.1 Sol is available in paid Work/Codex plans and the API; Terra adds a lower-priced tier. Rumored: an Astra-tier 6.1 refresh, with no announced date.

Release Date

December 2026 is the earliest rumored window; H1 2027 is the safer read. OpenAI has announced nothing.

Pricing

Per OpenAI's pricing page: Astra $10/$50, GPT-6.1 Sol $2/$10, GPT-6.1 Terra $1/$5, Luna $0.10/$0.50. The GPT-5.6 Sol promo is available 'at least through November 21, 2026' per OpenAI; no end date is announced, so treat 21 Nov as a floor and calendar-mark it.

Reports and sourcing

Staged October rollout to paid plans

OpenAI rollout reporting · High · 2026-10-16

Plus/Pro/Business/Enterprise rollout through October is reported by multiple outlets.

6.1 Astra refresh

Unnamed supply-chain / community reports · Low, rumor score 57 · 2026-10-16

No primary source. We are tracking it because OpenAI's 2026 cadence (5.5 → 5.6 → 6 within five months) makes a fast point release plausible.

Related reporting

Read the OpenAI family guide

Mistral Large 4 open weights: Everything We Know

Mistral · Released 29 October 2026 · Last updated 2026-11-02

Released: Released; earlier rumor score is historical.

The weights watch is resolved: Mistral announced the release on 29 October. We checked the public model manifest, configuration and license on 2 November. The historical rumor score records our earlier watch; the release itself is confirmed. The API docs still carry a Public Preview label.

Released features

  • Weights released 29 October, before the company deadline
  • 1.05T total / 52B active parameters in the model card
  • Text and image input; 1M context
  • Mistral Ecosystem Open-Weight License; separate Model-as-a-Service terms

What We Know

Mistral's release announcement and public Hugging Face manifest establish the weights release. The manifest lists 54 safetensors shards; configuration and license files are also public. This verifies the published artifacts, not a local inference benchmark or a full-model download.

Release Date

Mistral published the weights announcement on 29 October, meeting its month-end promise. This roundup is now a completed release record; no countdown to a future weights drop applies.

Pricing

The API docs list $0.70 input / $2.10 output per million tokens and $0.07 cached input. Weight availability is separate from the API's Public Preview lifecycle label. Review Mistral's license if you operate a Model-as-a-Service business.

Reports and sourcing

Weights promised by the end of the month

Mistral launch post, 6 Oct 2026 · High, on the record · 2026-10-06

'We will release the weights by the end of the month.' The post does not name a license.

No weights yet as of 27 Oct

Frontrumors check of Mistral's news page and Hugging Face · High, direct check · 2026-10-27

The newest mistralai repositories are the Timestral time-series models; no Large 4 repository appears among the most recently updated.

Weights released

Mistral release announcement · Confirmed, first-party announcement · 2026-10-29

Mistral released the improved Large 4 weights under the Mistral Ecosystem Open-Weight License. Its page publication metadata is dated 29 October.

Public model artifacts checked

Mistral's Hugging Face model repository · High, direct manifest and small-file check · 2026-11-02

Public, ungated repository; 54 safetensors weight shards, configuration and license listed. config.json returned HTTP 200; no full-model download or inference test was performed.

Related reporting

Read the Mistral family guide