By Assaf Biton
TL;DR: Between April and June 2026, OpenAI shipped GPT-5.5 ($5/$30 per 1M tokens), DeepSeek shipped the open-weight V4 (now $0.435/$0.87 for Pro), Google launched Gemini 3.5 Flash ($1.50/$9) at I/O, and Anthropic released Claude Fable 5 (then pulled it three days later under a US export-control order). Veo 4, Sora 3, Llama 5, and "Qwen 4" never shipped. Here is the real list as of June 19, 2026.
Everyone in my feed is hyped about models that do not exist. Veo 4, Sora 3, Llama 5, Gemini 4, "Qwen 4": all of them showed up in breathless threads this spring, and none are real. So I checked every frontier release from April through June 2026 against official pages and pricing, and threw out the rumors. What is left is short, strange, and useful.
Every release below is confirmed, current as of June 19, 2026, and priced from official sources. The fakes get a quick burial at the end, because knowing what is not shipping saves you just as much time.
The text models that actually shipped
Four frontier text models defined the quarter, and they could not be more different on price. OpenAI's GPT-5.5 arrived April 23 with a 1M token context window and benchmark wins on coding, at $5 per million input tokens and $30 per million output. One day later, DeepSeek dropped V4 (V4-Pro and V4-Flash), open weights under the MIT license, and made its 75% launch discount permanent on May 23: V4-Pro now costs $0.435 input and $0.87 output. That is roughly 34 times cheaper on output than GPT-5.5 for near-frontier performance.
OpenAI shipped GPT-5.5 at $5/$30 per 1M tokens with a 1M context window, then shut down the Sora video app and deprecated its API.
Best for: General knowledge workers needing a capable all-in-one assistant, Developers wanting quick code generation with GPT Image and Codex integration
Then Google answered at I/O on May 19 and 20. It launched Gemini 3.5 Flash at $1.50 input and $9 output, announced Gemini 3.5 Pro for a June rollout, and revealed Gemini Omni, a video-first model folded into the Gemini family.
Google launched Gemini 3.5 Flash ($1.50/$9) and Gemini Omni at I/O 2026; the rumored Veo 4 never shipped.
Best for: Google Workspace users wanting native AI inside Docs, Gmail, and Sheets, Researchers needing citation-backed outputs via NotebookLM
Want my honest read? For most paid work, GPT-5.5 and Gemini 3.5 still win on polish and ecosystem. But if your bill is the bottleneck, DeepSeek V4-Flash at $0.14 input is hard to argue with. Compare the math in our GPT-5.4 family pricing breakdown.
Pro Tip: Route cheap, high-volume calls (summaries, classification, drafts) to DeepSeek V4-Flash at $0.14/M input and reserve GPT-5.5 for the 5% of prompts that need its reasoning. On 10M tokens a day, that split can cut a $300 daily bill below $40.
The Claude Fable 5 saga: shipped, then pulled in 72 hours
This one is wild. Anthropic shipped Claude Opus 4.8 to general availability on May 28. Then on June 9 it released Claude Fable 5, the first public version of its restricted "Mythos" class, at $10 input and $50 output, alongside the locked-down Mythos 5 for government-adjacent security work.
Anthropic shipped Opus 4.8, then released and within days suspended Claude Fable 5 and Mythos 5 under a US export-control directive.
Best for: Knowledge workers doing deep document analysis and long-form writing, Developers who need precise instruction-following for complex multi-step tasks
Three days later it was gone. On June 12, a US export-control directive ordered Anthropic to cut off both models for any foreign national, anywhere, including its own employees. Unable to filter by nationality in real time, Anthropic pulled them for everyone the same day, June 12. Reporting calls it the first time a leading lab has taken a deployed model offline by federal order. The stated trigger: the government believed someone had found a way to jailbreak Fable 5.
If you wired anything to Fable 5 during its three-day window, it is offline with no restoration date. Fall back to Opus 4.8 ($5/$25), which stays available and covers nearly every non-security task.
Tooling moved quietly while the models made noise
The agent and coding-tool layer kept shipping without drama. Claude Code is still on the 2.1.x line, not the rumored 2.2: v2.1.179 landed June 16, and the patches kept coming, with v2.1.183 shipping June 19. The notable change came June 10 with v2.1.172, where sub-agents can now spawn their own sub-agents up to five levels deep. That is real recursion in a shipping CLI.
Still on the 2.1.x line (v2.1.183 as of June 19, 2026); sub-agents can now spawn their own sub-agents up to five levels deep.
Best for: Senior developers who want to delegate long autonomous tasks and review results, DevOps teams integrating AI into CI pipelines for automated test fixing
Around it, Google's Gemini CLI (Apache 2.0) and OpenAI's Codex sub-agent swarms kept the open-source pressure on, while OpenAI Workspace Agents and Anthropic Managed Agents traded blows in the enterprise lane. The plumbing matured too: A2A v1.0 went GA and MCP added OAuth 2.1, both now under Linux Foundation governance.
The rumors that never shipped
So which hyped releases are vapor? Quite a few, and the pattern is telling.
| Rumored release | What is actually true (June 2026) |
|---|---|
| Veo 4 | Never shipped. Google's I/O video reveal was Gemini Omni, not a standalone Veo 4. |
| Sora 3 | Does not exist. OpenAI killed the Sora app April 26; the API is deprecated and shuts down September 24, 2026. |
| Llama 5 | Not released. Meta's spring model was Muse Spark, a closed-weight pivot, not Llama. |
| Claude Sonnet 4.7 | Skipped. Anthropic's Sonnet line jumps to 4.8, still unreleased as of June 19, 2026. |
| "Qwen 4" and "Kling 4" | Wrong names. Alibaba ships Qwen 3.6; Kuaishou's current video model is Kling 3.0. |
Notice the theme: video models keep slipping, and loud "next major version" names tend to be fan inventions.
Pro Tip: Before you build on any model, search for its official model card and an API price per 1M tokens. If neither exists, it is not shippable yet. This one check killed every fake release above in under three minutes each.
Why this matters for what you pick
The real story of spring 2026 is not a single launch. It is a 34x output-price gap between the cheapest open-weight frontier model and the priciest closed one, plus a new risk: a model can now vanish overnight by government order.
If cost is your constraint, the open-weight tier (DeepSeek V4, Qwen 3.6, Kimi K2.6) is good enough for most jobs at a fraction of the price. If reliability matters more, the closed models still lead. But Fable 5 proved "available today" is not "available next week," so wire in a fallback model from day one.
Frequently asked questions
Is DeepSeek V4 really cheaper than GPT-5.5?
Yes, by a wide margin. As of June 2026, DeepSeek V4-Pro costs $0.435 input and $0.87 output per million tokens, while GPT-5.5 costs $5 input and $30 output. That makes V4-Pro roughly 34 times cheaper on output, and V4-Flash ($0.14/$0.28) is cheaper still. DeepSeek is also open weights under the MIT license.
What happened to Claude Fable 5?
Anthropic released Claude Fable 5 on June 9, 2026, then suspended it (and Mythos 5) on June 12 after a US government export-control directive that barred access for foreign nationals. As of June 19, 2026, both models are offline for all users with no restoration date announced.
Did Google release Veo 4 at I/O 2026?
No. Despite heavy speculation, Veo 4 was never announced. Google's I/O 2026 video reveal was Gemini Omni, a video-first model inside the Gemini family. Google's current shipping video models remain the Veo 3.1 line, including Veo 3.1 Lite.
What is the latest Claude Code version?
As of June 19, 2026, Claude Code is on v2.1.183, still in the 2.1.x series (v2.1.179 landed June 16, and v2.1.183 shipped June 19). There is no version 2.2. The standout recent feature, added in v2.1.172 on June 10, lets sub-agents spawn their own sub-agents up to five levels deep.
Our Take
Best value right now: DeepSeek V4 (open weights, MIT, $0.435/$0.87 for Pro) for high-volume work where you control the stack.
Best all-rounder: GPT-5.5 or Gemini 3.5 if you want polish, ecosystem, and support over raw price.
Your next step: pick one cheap model and one premium model, wire both into your app behind a single switch, and run your real workload through each for a day. The price gap is now big enough that the test pays for itself before lunch. Then bookmark this page, because by August half of it will be out of date.
