Is It Now an AI Cost-Efficiency War? Will OpenAI Turn the Tables with GPT-5.6?

Hello. I am reporter An Hye-min, who handles and works with data. OpenAI, which had been evaluated as falling behind Anthropic, has sharpened its knives and returned. As praise continues for its new model, GPT-5.6, the AI market is buzzing once again. In today's OhGraph, we prepared the story of GPT-5.6. Let's look through five graphs at why this model is drawing attention and how the AI market is shifting as a result.


Following Chess and Go, Now Coding... Top-Tier Programmers Fall to AI
Before diving into OpenAI, let's first go to Japan. There is a company called AtCoder, founded in 2012. It regularly hosts programming contests where programmers from around the world can showcase their skills. Among them is a tournament where only the top scorers from online qualifiers and regular competitions gather to compete: the World Tour Finals. It is an offline coding competition where the top 12 elite champions in AtCoder rating are invited to Tokyo to compete. This year, the tournament was held once again with top-tier programmers from all over the world.

Enlarge image

There are two main categories: Heuristics and Algorithms. Algorithm problems have a simple structure where you get points if you get the right answer and none if you get it wrong. In contrast, heuristic problems involve finding a "plausible best answer" based on human experience and intuition. They present problems where there is insufficient time to search through all possible cases, testing the process of deriving an efficient result.

Twelve programmers participate in each of these two divisions, and since last year, AI models have also been participating and competing. In last year's competition, AI models lost to human programmers in the heuristic category. Although AI's coding capabilities had improved, humans were still superior at the topmost level. The participant who defeated AI at the time was Polish programmer Psyho. How about this year?

Enlarge image

This year's heuristic problem was a stage lighting brightness puzzle. The task was to place mirrors, beam splitters, and absorption blocks on a grid stage measuring 20 cells wide by 20 cells long to reflect, split, and absorb light in various ways to achieve desired brightness levels. Instead of a single brightness level being given, there were 100 brightness levels to match. For instance, participants had to produce 100 randomly given brightness levels such as 0.214, 0.731, and 0.95. Since the maximum number of component changes on stage was limited to 200 times, altering the structure every time a brightness level was given was impossible. In other words, participants had to design a structure from the beginning that could efficiently represent multiple brightness levels to solve the problem. Who was the winner of this puzzle?

Enlarge image

In the heuristic algorithm problem conducted over two days, an AI model took an overwhelming first place. Although a human programmer briefly surpassed the AI model on the second day, the ultimate winner was the AI. The final score achieved by the AI was close to the maximum score of 50 billion points, a staggering sevenfold difference compared to the top human contestant. Thus, the heuristic division ended in a complete victory for AI.

Next is the algorithm division. A total of five problems were given in the algorithm division, with victory determined by who solved them correctly first. Looking at the line-up of programmers in the algorithm division, they are all truly remarkable figures in this field. There is tourist from Belarus, who competed in the International Olympiad in Informatics at the youngest age of 11 and set a record as the first in history to win six consecutive gold medals. There is also jiangly, a legendary Chinese programmer with a suspicious anime profile picture. Alongside them, representative programmers from the United States and Canada also participated to compete for supremacy; let's examine the results through a graph.

Enlarge image

In the early stage of the competition, hope emerged for humanity as Chinese contestant jiangly solved Problem A faster than the AI. But in the end, victory went to the AI once again. The AI solved all five problems, scoring a perfect 8,300 points. Not a single human programmer could solve Problems C and E. Notably, Problem E was not easy for the AI either; it was able to get the correct answer only after nearly 6 hours and 5 incorrect attempts.

Enlarge image

Both heuristics and algorithms ended in overwhelming victories for the AI model. Psyho, the last human who had defeated AI in the same competition last year, praised the AI's performance in dominating human competitors. Following Deep Blue defeating the chess champion in 1997 and AlphaGo defeating 9-dan Lee Sedol in 2016, AI has now defeated humanity in a top-tier coding competition. The AI model that defeated humanity was none other than OpenAI's model. An OpenAI researcher participating in the live broadcast explained that the model used in this competition was at a level similar to its new model, GPT-5.6.


OpenAI Bounces Back... Delivering Both Performance and Cost-Effectiveness
Despite this narrative, until recently, it was actually Anthropic's models that demonstrated outstanding performance in software development. Many developers were satisfied after using Claude Code, and companies scrambled to use Anthropic's products. Looking at actual enterprise LLM spending, Anthropic overtook OpenAI last year.

Enlarge image

In 2025, Anthropic accounted for 40% of enterprise LLM spending. Its market share, which stood at 12% in 2023, passed 24% to reach nearly half. Meanwhile, OpenAI lost nearly half of its market share, dropping to 27%. Looking at the coding market alone, Anthropic was dominant in 2025, capturing more than half of the market.

Conversely, OpenAI delivered results that fell short of expectations. For example, do you remember GPT-5 released last summer? Many people expected huge performance gains with the transition from 4 to 5, but when opened, reality proved different.

That does not mean OpenAI was standing still. It steadily pushed for performance improvements. Signs of change began to appear in the second half of last year. Let's head to Baku, the capital of Azerbaijan, in September 2025. It was the finals of the International Collegiate Programming Contest (ICPC) held in Baku. Roughly 3,000 universities participated, with only the top 139 teams advancing to the finals, where they had to solve 12 algorithm problems within a 5-hour time limit.

Enlarge image

First place went to Saint Petersburg State University, which solved 11 problems. It was followed by the University of Tokyo in second place, and Beijing Jiaotong University and Tsinghua University in third and fourth places. Gold medals were awarded up to fourth place. Notably, Seoul National University finished 10th, earning a bronze medal. OpenAI and Google DeepMind also participated in this contest. OpenAI solved all 12 problems, scoring higher than the university teams. Google DeepMind solved 10 out of the 12 problems. Incidentally, no university team was able to solve Problem C, but both AI models solved it.

OpenAI continued updating its versions following the release of GPT-5. And finally, GPT-5.6 emerged into the world. Of course, this model was not released to the market right away. The U.S. government placed restrictions, allowing only 20 partner organizations approved by the government to use it first. This is only the second time the government restricted access to an AI model in this manner, following Anthropic's Fable and Mythos 5. While accepting the government's request, OpenAI pointed out that such an approach should not become routine. On the other hand, it accepted regulation partly because it was facing an IPO, but also because it internally acknowledged the model's potential risks to a degree.

Enlarge image

OpenAI operates its own Preparedness Framework to assess model risk. It classifies capabilities under two benchmarks: High and Critical. High refers to a level where AI can significantly amplify existing threat vectors, while Critical refers to a level where AI can create entirely new threat vectors that would have been impossible without AI. This time, GPT-5.6 was classified as High in cybersecurity and chemical and biological domains. Consequently, OpenAI accepted the government's stance requesting a limited release.

After negotiations took place between OpenAI and the U.S. government, restrictions were lifted, and GPT-5.6 was officially released to the market. It was released across three tiers: Sol, the highest tier; Terra, a lightweight model; and Luna, an even lighter model. Metrics evaluating model capability show that in coding and finance, it reached the level of Anthropic's Claude Fable 5. Remarkably, in the case of the GPT-5.6 Sol Ultra model, it proved a math problem that had gone unsolved for 50 years in just 1 hour.

Along with performance, what people responded to was the model's cost-effectiveness. As new models emerge everywhere and performance becomes broadly leveled, cost is now what matters. Choices are being driven by which company manages token cost optimization. In terms of cost, GPT-5.6 is evaluated as highly competitive.

Enlarge image

The X-axis represents model cost, and the Y-axis represents model performance. Looking at models with similar performance levels, Anthropic's Claude Fable 5 is too expensive relative to its performance. But what about OpenAI's new model? While sitting on the same line in performance, its cost is far cheaper.


Anthropic Feeling the Heat? Fable 5 Free Trial Extended for a Third Time
Faced with this situation, Anthropic could not help but be anxious. Although it unveiled Fable 5 in early June, the service was suspended just three days later due to U.S. government control guidelines, preventing Anthropic from properly enjoying first-mover advantage. With controls lifted and service resuming only on July 1, nearly a month later, Anthropic might well hold a grudge against the U.S. government.

Enlarge image

Having released a model with top performance, Anthropic planned to use it to generate substantial revenue. It set up the new Fable 5 model not to be included in existing subscription plans, but rather billed separately based on usage. However, it announced that this policy would take effect later, and until then, allowed users to experience Fable 5 up to 50% of their subscription usage limits. The goal was to let many users try Fable 5, marvel at its performance, build interest, and later pay additional fees.

It was under these circumstances that OpenAI's GPT-5.6 was unveiled. With performance on par with Fable 5 and a lower price tag, consumers responded immediately, and Anthropic quickly took notice. The card Anthropic pulled out as a result was extending free trial benefits.

Enlarge image

Originally planning to end the trial on July 7 and transition Fable 5 to a paid model, Anthropic extended it once more so users could use it until July 12. Not stopping there, Anthropic extended it by another week, allowing Fable 5 usage until July 19. This shows just how much Anthropic is concerned about OpenAI's new model.

In truth, Anthropic is not unwilling to include Fable 5 in subscription plans. However, the reason for setting up a separate pricing model was insufficient computing resources. If included in subscription plans, Fable 5 usage would surge, and Anthropic does not yet possess the computing resources to handle that load. If that happened, latency would increase and servers would frequently crash, leading to a much worse user experience.

Enlarge image

To make matters worse, Elon Musk's xAI also released a new model. Having acquired Cursor, the top coding tool program since going public, Elon Musk unveiled Grok 4.5, a new AI model specialized for coding tasks. Elon Musk has explicitly targeted Anthropic's model "Opus" when discussing this model, asserting that their new model reaches Opus-level performance. Musk's key selling point here lies in cost-effectiveness: offering Opus-level performance while being faster, more efficient, and cheaper than Opus.

Enlarge image

The X-axis represents token usage, and the Y-axis represents model performance. Let's compare Grok 4.5 and Claude Opus 4.8 on the same line. Their performance is equal, but Grok 4.5 is overwhelmingly superior in token efficiency. The same goes for the previously mentioned GPT-5.6. While on the same line as Fable 5, it holds a far greater advantage in terms of cost-effectiveness. From the user's standpoint, models positioned in this quadrant are bound to be attractive.

Performance is rising, prices are falling, and the coding market once dominated by Anthropic is entering a new phase. Big tech companies' competition spreading beyond performance into price is not a bad thing for consumers. After all, it means better models can be used at lower costs.

However, the fact that most major AI companies remain unprofitable is a point to ponder. No one can guarantee how long this war of attrition can continue. Who will ultimately emerge as the winner of the frontier model race? That is all for today's OhGraph. Thank you very much for reading this long article to the end.

References
- Psyho(@FakePsyho) | X
- Ethan Knight(@__eknight__) | X
- AtCoder World Tour Finals 2026 Heuristic | AtCoder E
- nterprise LLM API Market Share by Usage | Menlo Ventures
- ICPC 2025 World Finals Baku | ICPC World Finals
- Preparedness Framework | OpenAI
- Artificial Analysis Coding Agent Index v1.1 | OpenAI
- A Proof Of The Cycle Double Cover Conejcture | OpenAI
- Prompt Used For "A Proof Of The Cycle Double Cover Conejcture" | OpenAI
- Vals Index Industry Average Accuracy Comparison | Vals AI

Written by An Hye-min | Design: Ahn Jun-seok | Intern: Shin Yeon-seong
※ Please note: This article was translated by AI and may contain errors.