Research
Model IndexMonthly10 min read

The Gamma Model Index: OpenAI takes the lead for the first time

November 2026. Revealing the model choices behind our agent as a window into the competition in AI for knowledge work.

For most of this year, Anthropic has dominated the competition for knowledge work, and here at Gamma we’ve consistently preferred their models for generating presentations and editing slides. But in the last few months, we’ve been happy to see some real competition at the frontier.

On October 29, we concluded three A/B tests that together set a record: for the first time since 2023, OpenAI surpassed Anthropic in our usage. Specifically, the new GPT-6.1 Terra replaced Claude Haiku 5.5 and Sonnet 5.5 for several key jobs in Gamma.

This post is a window into how we make those model decisions to serve millions of knowledge workers around the world. Gamma's agent handles millions of sessions each week, and we A/B test every new model on real users.

Our agent splits every deck into about a dozen jobs, and each one can run on a different model. Three come up throughout this post: the conversation, where the agent talks with the user and plans the deck; writing, which turns that plan into finished slides, layout and all; and vision, the jobs that look at rendered slides to review and repair them.

Presentations are an unusually good window into the AI race: coding agents tell you how models write software, and benchmarks tell you how they do at math, but decks are a business universal, and making one is a composite of many skills: research, writing, structure, data and design. Our users are executives and teachers, freelancers and large companies, in dozens of languages. If you want to know which models are winning the work regular people do, slides are the bellwether.

Fig. 1Each lab's share of our AI spend

All of our AI spend: our original generator as well as the new agent, and image models as well as language models. Trailing seven days at list prices, for weeks ending Sep 7 to Nov 6.

1September: OpenAI leads in vision

Through September, while we tested our new agent ahead of its October launch, OpenAI's share of our spend more than doubled, and the reason was volume. Every slide our agent writes is rendered and checked, and any slide that fails a check is sent back with a screenshot to be repaired. That's tens of millions of renders a week, each one needing a model that can look at a slide, understand what's wrong and fix it, cheaply. The Luna series dominated our testing here because of its strong vision, low cost, and appealing token latency.

The big models were still up for grabs. OpenAI's GPT-6 Astra and GPT-6 Sol both arrived in September with high expectations, but their time in the spotlight was brief: just days later, Astra lost our premium planning and writing to Claude Opus 5.5. GPT-6 Sol was cut from every test within a week, and GPT-6.1 Sol stuck around only for specialized jobs.

2Late September: Sonnet and Opus win on nuance

At the end of September, Claude Sonnet 5.5 arrived and won the most expensive job we have: talking with the user and planning their deck. The orchestrator is the job with the most context, the most turns and the most judgment, and Sonnet has defended it against every challenger since. We've always found that Anthropic's models are better at juggling several requirements at once, and at inferring the ones a user never spelled out, where GPT models tend to be too literal. Decks it planned were rated good about 18% more often than those of the strongest alternative. Around the same time, our premium tier moved entirely to Claude Opus 5.5, which writes the cleanest slides of any model we've tested — though it took significant prompting to veer it away from certain “AI slop” visual tells.

Anthropic's share stopped falling within days, and it held the lead for another month.

3October: Luna vs Haiku face off

October was a month of small models fighting over specific jobs, and the results split cleanly.

Claude Haiku 5.5 arrived on October 7 and took over editing decks on request. It cost 59% to 89% less per edit than the models it replaced, with the same quality, measured by how often people kept its edits. But Haiku lost every other job we gave it: planning, writing, reviewing and repair. The biggest challenge we faced was an inflexible thinking budget: “low” thinking wasn’t smart enough, and “medium” thought for far too long, with no good way to land in the middle.

Google's Gemini 3.8 Flash made a surprise re-entry to our leaderboard, after a summer of falling behind. It won the job of writing decks in our free tier, in part because it occupied a unique spot in the Pareto curve of speed, latency, and quality — above Luna/Haiku but below Sonnet/Sol. It still has major weaknesses, though: it overflowed slides about twice as often as OpenAI's GPT-6.1 Sol, which won writing for most of our other decks. In planning, where a model has to hold a conversation and call tools reliably, Flash lost every test, especially because Vertex reliability is far behind the other platforms.

4October 29: GPT-6.1 Terra competes on price

Then came GPT-6.1 Terra, which competed on price. It cost about half as much as the larger models before it, a price point no other model held: strong enough for hard jobs, and cheap enough that we could deploy a strong model at much wider scale than before.

Where it proved best was vision. On slide repair, it halved the share of slides still failing our checks after repair, from 9.1% to 4.7%. It cost more per deck than the small models it replaced, but it was worth it: ratings rose too. The same week it took over picking templates, at twice the speed and half the cost of the model before it. Repair is one of our biggest line items, and almost all of it now goes to OpenAI.

In the last full week of October, OpenAI took 40% of our AI spend, Anthropic 38% and Google 9%.

5Breaking down our inference bill

Looking at our model spending job by job, we’ve found that each lab owns a different kind of work.

Fig. 2Each lab owns a different kind of work

Each job's share of our agent's AI spend, images included, in the last full week of October, and how it splits among the three big labs; other providers make up the rest of each bar. Pick a job to see something it made.

Claude is the best orchestrator. Planning and talking with users is our largest line item, about 42% of spend, and 78% of it goes to Anthropic. Editing on request is another 9%, 71% of it to Anthropic. These are the jobs that need the most judgment about what a person intends, and we find that OpenAI models continue to struggle here. Claude consistently wins at “do what I mean, not what I say”.

OpenAI leads in vision and visuals. Repairing slides is 14% of spend, and 97% of it goes to OpenAI. Images are another 14%, and OpenAI paints most of them: GPT Image has gone from under 1% of our images in January to about half, beating Google's Nano Banana on cost-performance. People replaced images from GPT Image 2's fast setting about a third as often as those from Nano Banana 2's fast version, and it costs about a quarter as much per image as Nano Banana 2.1.

Writing is the open fight. It's 17% of spend, split three ways: 61% OpenAI, 30% Google, 10% Anthropic. It's also where a new release can move the most money fastest, and where we expect our own proprietary models to take the lead in the next few months.

6Volume vs. spend

The numbers above are based on cost. Count model calls instead, and the picture flips. Our agent made more than 70 million model calls last week, using more than 200 billion tokens, and OpenAI's two Luna models handled more than half of both. Claude Haiku 5.5 handled another fifth. Together they did three-quarters of the work for about a quarter of the language-model spend. Claude Sonnet 5.5 made one call in ten and took more than a third of the money.

Fig. 3Count model calls instead of spend, and the picture flips

Each model's share of our agent's model calls and of its language-model spend, Oct 27 to Nov 2.

The workhorses are small models, and they're where the competition is fiercest, because a model that wins a high-volume job takes millions of calls a day overnight.

7Why it moves so fast

Labs are competing for jobs. Application companies like Gamma compete with the labs by picking and choosing the best model for each job. We chose a model for each of about a dozen jobs, and re-chose whenever something new arrived. The result is a bill that can swing by a few points in a week. Claude Haiku 5.5 is the best editor we've tested and lost every other job. Terra transformed repair and lost at planning. A model that wins one job is worth real money, even if it loses the rest.

Fig. 4Jobs change hands whenever a better model arrives

Each of our agent's jobs, with its language-model spend split by model, week by week from Sep 14; our own models are left out. Hover or tap for a week's split.

At volume, vision has to be cheap. The jobs that look at rendered slides run on every slide, so they go to small models. OpenAI's small models can see a slide well enough to fix it. Anthropic's can't yet: asked to repair slides that failed our checks, Claude Haiku 5.5 left about 9% still failing, against 3.5% for GPT-6.1 Sol. Anthropic's large models see very well, but OpenAI’s do the same for much lower cost.

Newer isn't always better. GPT-6 Luna is cheaper per token than GPT-5.6 Luna, but the older model still plans our free users' decks, because it finished more of them, faster. And on two jobs we measured side by side, OpenAI's models wrote 2.6 to 4.6 times fewer output tokens than Claude's for the same work.

8Open models: strong on leaderboards, not yet at work

Open-weight models are having a moment. Some companies are moving their engineers' coding work onto them, and on public leaderboards for web design, DeepSeek's V4.1 Flash is the strongest open model there is. On our work, they haven't come close to their benchmarks.

DeepSeek V4.1 Flash lost our deck-writing comparison to GPT-6.1 Sol, 22 to 32, and to Claude Opus 5.5, 22 to 40. It spent so long reasoning that the first slide took about a minute to appear. With reasoning turned off, the first slide came in two seconds, but the design fell apart, down to the level of Gemini Flash. As a slide repairer it was 2.6 times faster than our default, and cost about three times as much per call. Z.ai's GLM did worse: the 30 decks it wrote in our benchmark had 29 serious failures between them.

Fig. 5Blind test: three models wrote the same deck. Which deck did the open model write?

One brief and one template from our offline writer bench, written by three models; images render as grey placeholders on this bench.

Open models matter to us anyway, as a starting point. Untuned, they lose. Trained on our data, against our checks, they become the specialists described below.

9What could move the bill next month

The biggest open question is the free tier's conversation. Since October 28, half of our free users' decks have been planned by Terra and half by Sonnet, and so far the results are close. If Terra wins, OpenAI takes a share of Anthropic's biggest job. If it loses, the conversation stays Claude's.

We're also testing Opus 5.5 at higher thinking for our premium tier, a three-way review test between Sonnet, Terra and Sol, and Opus against GPT-6 Luna for writing custom themes. Google's Gemini 4 Argon was unreliable in our first trial, with a fifth to a third of its calls timing out, but we're now starting to see promising results from Gemini 4 Hydrogen. All eyes are on Anthropic's new Claude Fable 5.2, released November 9, but most of our volume runs on small models, so the race that matters most to us is small models with good eyes.

And there's a line that isn't on the chart yet: our own. Our research team, built around our acquisition of Lica, has trained a slide-repair model that fixed 76% of failing slides on the first try, against 57% for the general model it was tested against. We expect our own models to be the fastest-growing line in this index in the next few months, and they represent another risk factor for the big labs: a world where token volume continues to rise, but competition forces costs to come down. That future is a win for consumers, and we’re doing our best to accelerate it.

10The scorecard

Fig. 6Each model's verdict, from the jobs it won and lost

Open a row for the model's share of each job and the evidence behind its verdict.

These verdicts come from randomized experiments on real users, judged by what they did with the result (whether they kept the AI's edits, presented the deck, shared it, exported it, and paid), and from our offline comparisons, where we've created thousands of evals to compare decks side by side. Every slide also gets automatic quality checks, and we track speed and cost. Spend is measured at list prices, across the language and image models in our agent. We ran more than twenty of these experiments in October. See you in December.

Jon Noronha is a co-founder of Gamma.

@article{gamma2026modelindex, title = {The Gamma Model Index: OpenAI takes the lead for the first time}, author = {Noronha, Jon}, journal = {Gamma Research}, year = {2026}, month = {nov}, url = {https://research.gamma.app/2026/11/model-index} }
More from Gamma Research
Getting frontier performance without frontier models
Models · Nov 2026