GPT-6.2 Astra and Sol are out. Claude Fable 5.5 and Opus 5.6 are still speculation, and a reported Gemini 4 Neon billing entry does not mean the model launched. In this World of AI roundup, the most useful news is less dramatic than the rumors: OpenAI has a cheaper Astra, Google is putting background tasks in more people's hands, and Microsoft has new coding and image models to test against the ones you already use.
Watch the World of AI Roundup
Episode sponsor: Browserbase via the creator's sponsor link. Its documentation describes cloud browsers for automation. The link is included for context, not as a JQ AI SYSTEMS recommendation.
What Shipped and What Is Still a Rumor
| Story | Status on 18 November 2026 |
|---|---|
| GPT-6.2 Astra and Sol | Released. OpenAI published model details and prices. |
| Silicon 2 / Fable 5.5 / Opus 5.6 | Reported sighting, unverified identity. Anthropic's public model lineup still names Fable 5.2 and Opus 5.5. |
| Gemini 4 Neon and Krypton | Unconfirmed. Neither appeared in the public Gemini API model catalog we checked. |
| Tasks in Gemini | Rolling out free of charge. Google lists six countries; some features have narrower eligibility. |
| Microsoft MAI Code 2 and Image 2.7 | Announced at Ignite. Access varies by model and platform. |
GPT-6.2 Astra and Sol: Better Value, Mixed Creative Results
OpenAI positions Astra for harder projects and Sol for everyday work. At standard short-context API rates per million tokens, Astra costs $8 input, $0.40 cached input and $40 output. Its input and output rates are 20% lower than GPT-6 Astra's. Sol keeps GPT-6.1 Sol's rates: $2 input, $0.10 cached input and $10 output. Those prices do not tell you the bill for an entire run; token counts, cache hits, tool calls and retries matter.
OpenAI reports that Astra gained 5.2 points over its predecessor on Agents' Last Exam at less than a third of the cost. That is a result on a specific suite of professional tasks, not a prediction for a website or game. World of AI's creative demos show why: Astra produced a more convincing dungeon-game prototype after the host raised reasoning effort. The Three.js ship demo was faster and cheaper than Fable's in the example shown, but the host preferred Fable's visual quality. The host also liked Sol's value in a comparison with Opus, while finding the new GPT models' frontend design taste inconsistent. These are first impressions, not a controlled benchmark or a measure of community consensus.
Pat Simmons's separate three-build test gives that trade-off some numbers: Astra 6.2 was the cheapest on all three runs, but Fable ranked first on each and Astra's baseball game finished last. For another perspective, the OpenAI demo below shows Astra rebuilding a set in Blender from two video frames. It illustrates the workflow rather than settling the model comparison.
Silicon 2 Is a Reported Label. Code Mode Is Real.
A community post reports seeing Silicon 2 in a Devin Build model list. The episode calls the displayed $10 input and $50 output rates placeholders. They match historical Fable 5.1 pricing, but Anthropic's own Fable page lists Fable 5.2 at $8 input and $40 output per million. Neither the label nor the price identifies a model. Separately, METR's November 5 assessment uses Model 4 as a label for an internal Anthropic model. It does not identify it as Silicon 2, Fable 5.5 or Opus 5.6.
Claude Code's experimental Code Mode is available to try and off by default. It lets Claude group related tool calls in JavaScript, pass results between calls and return just the output needed for the next decision. It does not bypass permission rules or approval prompts. A sensible trial is a repository task you already understand, with the same tests and permissions as a normal session.
Linked billing is a separate Claude Platform beta. An eligible customer can link their Claude organization to a third-party app, set a spend cap, and unlink later. The app specifies on supported API requests who pays. That is different from handing the app an organization API key or moving every charge to the customer.
Gemini: Watch Neon and Krypton; Try Tasks Where Available
Neon reportedly appeared in Google's Cloud billing catalog at $3 input and $15 output per million tokens. The sighting is a clue about a possible billable service, not evidence that anyone can call a public endpoint or that it launched on the date predicted in the episode. The Krypton showcase has a different problem: output attributed to an unidentified checkpoint does not establish which model made it or how reliably it outperforms Carbon or Fable.
What Google does document is Tasks in Gemini on Gemini 4 Hydrogen, rolling out free of charge in the U.S., India, Canada, Australia, Japan and South Korea. Start with a recurring research digest or a calendar briefing. Some actions, including Google Pay checkout, have separate eligibility; Google documents explicit approval before purchase. For readers in Portugal, the free rollout announcement does not list Portugal.
Two practical developer updates are easier to verify: Google AI Studio announced custom domains for Build projects, and the Gemini API changelog documents direct Office/OpenDocument inputs and selecting particular pages, cells or media intervals. Selecting only the material a task needs is a concrete way to reduce what an agent processes. A missing visible thinking trace in a Carbon demo, by contrast, doesn't establish that the model skipped reasoning.
Microsoft MAI and Kling: Test the Actual Output
MAI Code 2 and Code 2 Flash target everyday engineering and high-volume routine tasks, respectively. Microsoft's Ignite comparison puts Code 2 at 98% of GPT-6.1 Sol's score at 40% of the price, averaged over four selected coding benchmarks. That is a vendor result against 6.1 Sol, not a head-to-head with 6.2 Sol. Try Code 2 on a scoped repo change and Flash on a small fix before extrapolating to longer jobs.
MAI Image 2.7 and Flash are in Microsoft Foundry public preview at roughly $35 and $10 per 1,000 images, respectively, under Microsoft's examples. Kling Image 4.0 advertises native 8K output and coordinated image sets. For a production comparison, try editing the same product shot on each model: check that the logo and packaging text stay intact, then count the generations it took to get a usable result.
Other Stories Worth Tracking
- Devin Bots: Cognition's announcement describes persistent agents that coordinate Devin sessions around ongoing goals. Review work in bounded steps and keep merge and production permissions separate during a pilot.
- Grok 4.8: Musk described a limited rollout and a wider-release estimate. Treat the window as an estimate, not a date to build around.
- NVIDIA: The 17 November earnings release identifies Colette Kress as interim CEO and CFO. NVIDIA reported $115.8 billion in quarterly revenue, including $107.6 billion from Data Center. The disclosed figures are a firmer basis for tracking the infrastructure story than personal allegations.
- Perplexity: CEO Aravind Srinivas said the company crossed $1 billion in ARR. That is a company claim, not audited annual revenue or proof of profitability.
- Tesla Powercore: Tesla announced a solid-state medium-voltage transformer with inverter and breaker functions, targeting operation in 2027. It manages and converts power; it does not generate it.
What to Test First
Pick one task you already pay for in time or tokens. Run Sol and your current default with the same prompt, tools and acceptance checks. Use Astra or Fable for the hardest case that still fails. For code, measure tests passed and minutes of human review. For a frontend, also check narrow-screen layout and whether the key interaction works. Report cost per accepted result, including retries, alongside elapsed time. That's a better basis for switching models than a demo or the latest leaderboard position.
X Accounts and Source Posts
These are creator and community accounts from the video description. Their timelines show ongoing discussion, not confirmed release notes. Each profile also has a direct link if the embedded feed does not load.
Original Posts Behind Three Headlines
Video Chapters
Claude and Silicon 2
| Time | Topic |
|---|---|
| 00:00 | Intro |
| 01:51 | Fable 5.5 or Opus 5.6 speculation |
| 02:15 | Reported placeholder pricing |
| 03:00 | METR Model 4 |
| 05:03 | Fable checkpoint speculation |
GPT-6.2 Astra and Sol
| Time | Topic |
|---|---|
| 05:49 | Astra and Sol launch |
| 06:12 | Community reaction |
| 07:45 | What improved |
| 08:10 | Launch framing |
| 08:31 | Testing advice |
| 08:55 | Pricing |
| 09:20 | Benchmarks |
| 09:41 | Leaderboard snapshot |
| 10:30 | Context limits |
| 11:41 | Concerns |
| 12:21 | Verdict |
| 14:16 | The reported bel checkpoint |
| 15:10 | Capabilities |
| 16:15 | OpenAI model progression |
| 17:18 | Astra vs. Fable |
| 18:35 | SVG demo |
| 18:53 | Minecraft Dungeons-style build |
| 19:44 | Sol vs. Opus |
| 20:02 | Blender test |
| 20:30 | Three.js ship demo |
| 21:04 | Frontend design |
| 21:23 | Computer use |
| 21:40 | Pros and cons |
| 22:15 | Claude comparison verdict |
Gemini and the creative tools
| Time | Topic |
|---|---|
| 23:00 | Krypton speculation |
| 23:30 | Output attributed to Krypton |
| 25:33 | Neon billing-catalog report |
| 26:15 | Carbon thinking reports |
| 27:00 | Tasks in Gemini |
| 28:00 | AI Studio updates |
| 28:37 | Project Astra |
| 29:11 | Antigravity usage limits |
| 29:28 | Kling Image 4.0 |
| 29:50 | Kling vs. Nano Banana |
| 30:10 | MAI Code 2 |
| 30:52 | MAI Image 2.7 |
Agents, business and infrastructure
| Time | Topic |
|---|---|
| 31:27 | Claude Code Mode |
| 32:10 | Linked billing beta |
| 32:39 | Claude Desktop |
| 33:07 | Grok 4.8 timing |
| 33:35 | NVIDIA leadership |
| 34:27 | NVIDIA earnings |
| 35:24 | Demis Hassabis on long-term medical research |
| 36:12 | Devin Bots |
| 36:42 | Perplexity revenue claim |
| 37:33 | Tesla Powercore |
| 38:57 | Outro |
Sources and Evidence Cutoff
Checked 18 November 2026. The embedded episode and supplied transcript are the sources for World of AI's commentary and demos. Product announcements and documentation are linked beside the release claims. Social posts support attributions, not anonymous model identities or launch dates. JQ AI SYSTEMS did not rerun the demos or independently benchmark the models.