What Six Months of Model Releases Say About Enterprise AI
By Zeev Grinberg, Head of GenAI at Ness Technologies
Every major model release since ChatGPT is logged in this site's release tracker. That gives me something most commentary lacks: a dataset instead of a feeling. Looking at the first half of 2026 as one table, four patterns stand out, and each one changes a decision enterprises are making right now.
I run the GenAI practice at Ness Technologies, so I read these releases the way our clients have to: not as launch news, but as procurement, architecture, and risk questions. Here is what the data says.
The price floor keeps falling, and the ceiling stopped mattering
In February, Gemini 3.1 Pro arrived at $2 in and $12 out per million tokens while topping 13 of 16 benchmarks. In April, Grok 4.3 delivered frontier quality at $1.25 and $2.50. DeepSeek V4 Pro brought a 1.6 trillion parameter open model, MIT licensed, at $1.74 and $3.48. By July, Grok 4.5 launched at $2 and $6, and Anthropic priced Sonnet 5 at $2 and $10. The only releases holding premium pricing are the frontier tiers, GPT-5.5 at $5 and $30 and Claude Fable 5 at $10 and $50, and those are bought for hard problems, not for volume.
The practical consequence: for most enterprise workloads, the cost of intelligence is no longer the budget line that decides the project. Integration, evaluation, and change management now cost more than the tokens. If your business case for an AI project died on API pricing in 2024, it deserves a rerun in 2026.
List prices at launch, from the release tracker. Claude Fable 5 ($50, frontier tier) and intro discounts excluded.
The context window race is over. Token efficiency is the new race
A year ago a million tokens of context was a headline. In 2026 it ships as a default: GPT-5.5, Grok 4.3, DeepSeek V4 Pro and Claude Fable 5 all run at or near 1M, and Grok 4.5 chose 500K on purpose. When everyone has the same number, the number stops being the story. The new competition is what a model does with the tokens. xAI's headline claim for Grok 4.5 was not a benchmark score, it was efficiency: 4.2 times fewer output tokens than Opus 4.8 on SWE-Bench Pro.
For architects this shifts the question from how much can I put into the prompt to what each token of output costs at scale. Retrieval is not dead, it is just no longer a workaround for small windows. It is a cost and latency decision.
Nobody launched a chatbot this year
Read the release notes from the first half of 2026 and one word repeats: agentic. Grok 4.5 was built for coding and agentic work. Sonnet 5 is Anthropic's most agentic model and now the default for free users. Microsoft's first in-house family, MAI-Thinking-1, is a reasoning line. OpenAI's July release was not a bigger text model at all, it was GPT-Live-1, full duplex voice that listens and speaks at the same time.
The vendors have collectively decided that the next interface is a colleague, not a chat box. Enterprises should read that as a signal about where the platforms will invest their tooling, security models, and pricing over the next two years. Pilots designed around a chat window are pilots designed for 2024.
The alliances got strange, and vendor risk became a board topic
Apple rebuilt Siri on Google's Gemini models. Microsoft shipped its own models to reduce dependence on OpenAI while remaining its largest backer. xAI now releases under the SpaceXAI brand after the SpaceX merger. And the strongest open models come from DeepSeek and Alibaba, which for many enterprises adds a regulatory dimension to an engineering choice.
The lesson I give clients has not changed since 2024, but the evidence keeps getting stronger: build model-agnostic. The leader changed three times in six months by my own leaderboard. Any architecture that hardwires one vendor is a bet the vendors themselves are not making.
What I would do with this in Q3
Three concrete moves. First, re-price the use cases you shelved in 2024 and 2025 against current per-token costs; several of them are now viable. Second, add token efficiency and output verbosity to your model evaluations, not just quality scores, because at volume that is where the money goes. Third, if you are still piloting chat assistants, redirect at least one pilot to a bounded agentic workflow with human approval gates, because that is the pattern every major vendor is now building for.
The full dataset behind this piece is public: the release tracker and the live model comparison on this site. I publish this analysis monthly. If you want it in your inbox, the weekly newsletter covers what changes in between.