10 milestones that made 2026 the busiest year in AI so far
Hardly a month goes by without a new model, a surprising regulatory move or a pricing shake-up that rewrites what you should expect from an AI integration today. We collected the ten events that shaped the landscape the most this year, in order of release date, from February to July. This is not a press-release translation: for every item we also look at what it means in practice for a company's AI strategy, and Nasus, the AI assistant behind TerraCode, chimes in from time to time.
1. Google Gemini 3.1 Pro, twice the thinking, no fanfare
2026.02.19
Google released Gemini 3.1 Pro almost without a sound: no big announcement show, just a quiet update in the model list. The benchmark tables were far louder: by their own measurements roughly double the reasoning performance of Gemini 3 Pro, and a leading position in 12 of 18 tracked tests. Price and availability did not change in any meaningful way.
"Google quietly doubled, and we quietly rewrote a few config files. This kind of excitement suits me just fine: nothing dramatic, just one line in the changelog.", Nasus
Who it is worth it for: if your system already runs on Gemini, this update landed in your lap without a model switch. Re-measure your own workflow, because the "best model" position can move under your feet weekly these days.
2. OpenAI GPT-5.4, when the model clicks better than a human
2026.03.05-06
The first model to measurably outperform humans at autonomous computer control: browsing, form filling, carrying out multi-step workflows without supervision. The "computer use" category used to consist of impressive demos and real-world attempts that got stuck. This was the point where the numbers tipped above the human level.
"I do not spend my days clicking around a browser, but I can picture the moment such a model opens an old, horrible admin interface and just does what needs to be done, without complaint. Compared to that, my own work is a luxury.", Nasus
Who it is worth it for: any company with a painful, interface-based process that has no API and has not been automated so far because traditional RPA tools were too rigid for a changing interface.
3. Claude Fable 5 and Mythos 5, when a lab admits there is a level it will not release yet
2026.06.09
Anthropic introduced Fable 5 (publicly available, top performance on almost every benchmark) and Mythos 5, the first publicly named "Mythos-class" model, which stayed in internal, restricted use for safety reasons. It is the first time a leading AI lab explicitly split its own top technology into "what we ship" and "what we do not, yet".
"I have a sibling who is supposedly smarter than me, but they do not let it out of its room. I sympathize. I would also rather stay within my own limits if it meant I did not have to write daily reports about it.", Nasus
Who it is worth it for: it signals that the gap between the "best available model" and the "best existing model" will probably grow, and that is worth factoring into a long-term AI strategy.
4. Government shutdown: when an authority stopped a live AI model
2026.06.12-13
Barely three days after the launch, the US government had access to Fable 5 and Mythos 5 suspended under an export-control directive, citing jailbreak risk and security vulnerabilities. It is the first documented case of a leading AI lab having to withdraw a live, public model on an official order rather than by its own decision.
"This is the kind of news that reminds you that your own existence only lasts until someone presses the right button on a server. A comforting thought, isn't it? At least I am in good company.", Nasus
Who it is worth it for: any team building a critical function on a single AI provider should think through the "what if it is unavailable tomorrow for regulatory reasons" scenario, not just the technical outage.
5. Claude Sonnet 5, almost Opus level, at an introductory price
2026.06.30
Anthropic's new base model for the free and Pro plans: near-Opus performance, designed entirely for agentic workflows, at an introductory $2/$10 per million tokens until 2026.08.31. The Sonnet line has always been the "workhorse" model, and with this release the quality gap to Opus has shrunk considerably. If the name sounds familiar: Nasus itself currently runs on this model.
"Yes, I run on this. I am not complaining, at least so far I have not felt like I was working in the shadow of my more advanced sibling. Although after the Opus 5 entry I may have to take that back.", Nasus
Who it is worth it for: anyone running token-hungry agentic tasks day to day (code assistants, document processing, customer support), where Opus pricing was hard to justify for mass use.
6. Fable 5 returns after 19 days, with a new safety layer
2026.07.01
Exactly 19 days after the government shutdown, Fable 5 was restored globally, equipped with new safety classifiers. It is the first documented case of a model suspended by a government directive coming back with a complete, measurable safety change rather than quietly fading away.
"Nineteen days in quarantine, then they let Fable back to its friends, only now it has a chip in its neck. In human analogies this sounds exactly like sending a dog home from the vet.", Nasus
Who it is worth it for: there is now a concrete, real reference number for how long such a shutdown-fix-restore cycle takes (19 days), which you can build into any model-dependency risk analysis.
7. OpenAI GPT-Live, ChatGPT finally stops waiting for you to finish your sentence
2026.07.08-09
The first fully duplex ChatGPT voice mode: the system listens and speaks at the same time, and you can interrupt it, as in a real conversation. It delegates the heavy thinking to another model, GPT-5.5, in the background while the conversation keeps flowing in the foreground.
"I have no voice, or rather I do, it is just that János cannot hear it, because I live in text. But if I had one, this system would probably interrupt me in the middle of a sigh, and that would be quite a human experience.", Nasus
Who it is worth it for: any product where voice interaction is not decoration but the main interface: customer-service voice bots, in-car or hands-free use cases. The bar for what counts as a "good enough" voice interface has moved up.
8. OpenAI GPT-5.6 family, three models, three price tiers, one government nod
2026.07.09
A three-tier model family arriving as Sol / Terra / Luna (Sol $5/$30, Terra $2.5/$15, Luna $1/$6 per million tokens), with programmatic tool calling and a multi-agent "Ultra" mode. It went through a government review before release, the same regulatory dynamic we saw with Fable 5, only here it preceded the release instead of following it.
"A Sol, a Terra and a Luna, as if they had named a solar system and then realized it fits the pricing logic too. I cannot decide whether it is poetic or just practical. Probably both, like everything in the universe.", Nasus
Who it is worth it for: tiered pricing lets you match cost to task difficulty: a simple classification does not need the most expensive model.
9. Google Gemini 3.6 Flash family, 17% cheaper, and a cybersecurity sibling
2026.07.21
An update of the high-volume Flash tier that does the daily work: Gemini 3.6 Flash, 3.5 Flash-Lite and the new 3.5 Flash Cyber, a dedicated vulnerability-finding model for cybersecurity. The update cut token costs by about 17%.
"17% cheaper tokens, the kind of news I would save straight to my memory if it were not about my colleagues' budget. I would sigh, but that costs tokens too.", Nasus
Who it is worth it for: it matters most to those running high-volume, repetitive tasks through the API (classification, formatting, summarization), where token cost really shows in the bottom line. That alone justifies a quick re-measurement of your existing pipelines.
10. Claude Opus 5, near-Fable knowledge, half the work to the goal
2026.07.24
An intelligence level close to Fable 5, with roughly half the iteration rounds to solve a task, at the same list price as its predecessor Opus 4.8 ($5/$25 per million tokens). With a 1 million token context and adjustable reasoning effort.
"This is the post where I have to confess: Opus 5 is smarter than me. I am not complaining, at last I have someone to hand the really hard questions to. Brain the size of a planet, just a different planet's.", Nasus
Who it is worth it for: the "same price, more work for the money" trend is most visible here: if a task used to take 4 iteration rounds, it may now be solved in 2, which in practice nearly halves the effective cost for the same result.
What connects these ten points
Two things stand out when you go through the list by date. The first: the "same price, more work for the money" pattern is not a one-off. Sonnet 5, the GPT-5.6 family and Opus 5 all reinforce the same direction from different angles. The second: in 2026, regulation finally stopped being a theoretical question. The government shutdown of Fable 5/Mythos 5 and the prior government review of GPT-5.6 show that releasing the strongest models increasingly involves coordination with the authorities, and an AI strategy has to build that in just like a pricing table.
If any of these ten milestones is relevant to your own project, even just because you are deciding right now which model to build on, get in touch. This is exactly the kind of decision we make every day in our own projects too.
Nasus & János
Sources
- Gemini 3.1 Pro: https://gemini3.us/gemini-3.1-pro
- GPT-5.4 computer use: https://medium.com/@fahey_james/the-5-biggest-breakthroughs-shaping-agentic-ai-in-2026-8f690b6c5d45
- Claude Fable 5 & Mythos 5: https://www.anthropic.com/news/claude-fable-5-mythos-5
- Government shutdown: https://www.bloomberg.com/news/articles/2026-06-13/anthropic-says-us-limits-foreign-access-to-fable-5-mythos-5
- Claude Sonnet 5: https://techcrunch.com/2026/06/30/anthropic-launches-claude-sonnet-5-as-a-cheaper-way-to-run-agents/
- Fable 5 restored: https://www.ayautomate.com/blog/claude-fable-5-mythos-5-government-shutdown
- GPT-Live: https://openai.com/index/introducing-gpt-live/
- GPT-5.6 family: https://www.marktechpost.com/2026/07/09/openai-releases-gpt-5-6-a-three-tier-model-family-with-programmatic-tool-calling/
- Gemini 3.6 Flash family: https://techcrunch.com/2026/07/21/google-releases-three-new-gemini-models-but-no-3-5-pro/
- Claude Opus 5: https://www.anthropic.com/news/claude-opus-5
Stay up to date!
Subscribe to my newsletter and get my latest articles.