Latest / Neural Newscast / OpenAI Slashes GPT-5.6 Prices as Fields Medalist [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast, this is Model Behavior, AI-focused news and analysis on the models [00:05] Announcer: shaping our world. [00:11] Nina Park: Welcome to Model Behavior. [00:14] Nina Park: In this program, we provide a forensic examination of how artificial intelligence systems are [00:20] Nina Park: being constructed, deployed, and utilized within professional environments. [00:25] Nina Park: We focus on the intersection of technical innovation and the regulatory and economic [00:30] Nina Park: frameworks that define the current AI landscape. [00:32] Nina Park: I am glad you could join us for this briefing. [00:35] Thatcher Collins: Today is July 31, 2026, and we are monitoring a major inflection point in the economics of [00:42] Thatcher Collins: frontier models. Yesterday's comprehensive pricing announcement from OpenAI has effectively [00:47] Thatcher Collins: shifted the cost structure for the entire industry. It appears we are entering a new [00:52] Thatcher Collins: phase where the price of intelligence is falling faster than most analysts predicted. [00:57] Nina Park: OpenAI has effectively ignited a new phase of the AI price wars. [01:02] Nina Park: They've announced substantial cuts to their GPT-5.6 series, with the smallest model in [01:07] Nina Park: that family, Luna, seeing its price slashed by 80%. [01:11] Nina Park: This moves Luna into a much more competitive tier, pricing it at $0.20 per million input [01:16] Nina Park: tokens and $1.20 per million output tokens, Thatcher. [01:21] Thatcher Collins: It is a massive move, Nina. [01:23] Thatcher Collins: For context, that puts Luna's combined price at $1.40 per million tokens. [01:28] Thatcher Collins: Looking at the broader market, that undercuts Google's Gemini 3.5 Flashlight, which sits at a combined $2.80. [01:35] Thatcher Collins: It seems OpenAI is no longer content just competing on intelligence. [01:39] Thatcher Collins: They are now aggressively targeting the low-cost inference tier to dominate developer volume. [01:45] Nina Park: Exactly. And it was not just Luna that saw a shift. [01:49] Nina Park: The mid-tier model, GPT 5.6 Terra, saw a 20% reduction, bringing it down to $14 per million tokens total. [01:58] Nina Park: This aligns it exactly with the pricing for Google's Gemini 3.1 Pro Preview for context windows under 200,000 tokens. [02:05] Nina Park: However, the flagship model, Sol, is taking a different path entirely regarding its valuation. [02:11] Nina Park: You are referring to the new Sol Fast mode? [02:14] Thatcher Collins: I find that interesting because it actually goes in the opposite direction regarding cost. [02:18] Thatcher Collins: OpenAI is adding a premium mode that delivers two and a half times the throughput for twice the price. [02:24] Thatcher Collins: At $70 per million tokens combined, Solfast is now one of the most expensive configurations on the market. [02:31] Thatcher Collins: Nina, is there really a demand for that? [02:33] Nina Park: For enterprise applications that are latency sensitive, the answer is often yes. [02:40] Thatcher Collins: Think about real-time customer support or interactive coding assistance, where milliseconds directly impact user retention. [02:48] Thatcher Collins: By keeping the sole standard price at $35 but offering a high-speed alternative, they [02:55] Thatcher Collins: are essentially segmenting the market between those who need raw power and those who need [03:00] Thatcher Collins: extreme response speed for specialized workflows. [03:04] SPEAKER_03: It also seems like a defensive move against Anthropic. [03:07] SPEAKER_03: Claude Opus 5 was recently released at the same price point as its predecessor, effectively [03:12] SPEAKER_03: lowering its price per unit of capability. [03:15] SPEAKER_03: By slashing Luna and Terra, OpenAI is forcing the industry to look at the price per intelligence [03:21] SPEAKER_03: metric rather than just the sticker price of a single token. This forces competitors [03:26] SPEAKER_03: to justify their pricing against very efficient baselines. [03:30] Thatcher Collins: VentureBeat actually highlighted this, noting that OpenAI is now selling what was [03:34] Thatcher Collins: flagship intelligence just four months ago for about 1 13th of the price. [03:41] Thatcher Collins: It is a rapid commoditization of reasoning capability, [03:44] Thatcher Collins: But while the models are getting cheaper to run, the people building them are becoming even more specialized. [03:51] Thatcher Collins: That leads us to the other big news from today, Jacob Timmerman joining OpenAI. [03:55] SPEAKER_03: This really caught the academic community by surprise, Nina. [03:59] SPEAKER_03: Timmerman is a Fields Medal winner, an honor often described as the Nobel Prize of Mathematics. [04:04] SPEAKER_03: He is taking a leave from the University of Toronto to work specifically on AI safety, [04:10] SPEAKER_03: One machine-learning professor compared it to hiring Lionel Messi as a project manager. [04:15] SPEAKER_03: Why would someone at the peak of pure mathematics pivot to an AI lab right now? [04:19] Thatcher Collins: According to an interview with The Atlantic this morning, [04:22] Thatcher Collins: Sherman believes we are at a turning point where AI systems are beginning to solve research-level math problems. [04:29] Thatcher Collins: He pointed to how OpenAI models recently disproved an 80-year-old conjecture [04:34] Nina Park: and how Anthropix Claude resolved an even older one, [04:38] Nina Park: He sees math as the necessary language to take our fuzzy intuitions about AI safety [04:43] Nina Park: and make them technically precise. [04:45] SPEAKER_03: But Nina, he is coming from number theory, not computer science. [04:51] SPEAKER_03: In the interview, Zimmerman admitted he lacks a background in software engineering. [04:56] SPEAKER_03: He argued that current AI systems are now so good at coding [05:00] SPEAKER_03: that his lack of programming training is no longer an impediment to running large-scale experiments. [05:06] SPEAKER_03: That is a stark statement about the current state of productivity tools and their impact on research. [05:13] Nina Park: It is, Thatcher, but I have to push back a bit on the safety focus. [05:18] Nina Park: Many skeptics argue that these high-profile hires are often just more capability researchers in disguise. [05:25] Thatcher Collins: If Zimmerman helps these models reason more effectively through mathematical poofs, [05:30] Thatcher Collins: isn't he also making them more capable of hacking or bypassing existing safeguards? [05:35] Thatcher Collins: it seems like a dual-use problem that mathematics might not easily solve. [05:40] SPEAKER_03: That is the dilemma we always talk about. [05:43] SPEAKER_03: Zimmerman specifically mentioned a report he co-wrote last year on omnicide scenarios, [05:49] SPEAKER_03: where he mapped out how AI could lead to catastrophic outcomes. [05:53] SPEAKER_03: He says his focus is strictly on the safety side, [05:57] SPEAKER_03: particularly on developing a theoretical understanding of why neural networks behave the way they do. [06:03] SPEAKER_03: He argues that without precise definitions, we're just guessing at control. [06:09] Nina Park: He also expressed support for an open letter signed by over a thousand employees at top AI labs [06:15] Nina Park: calling for the United States government to coordinate a slowdown if development reaches a certain danger threshold. [06:22] Thatcher Collins: It is a rare moment of a top-tier academic entering the industry with a very public and cautious stance on international coordination. [06:30] Thatcher Collins: It will be interesting to see how he influences internal policy. [06:34] SPEAKER_03: And while OpenAI is securing the talent, Anthropic is securing the physical infrastructure. [06:40] SPEAKER_03: CNBC confirmed today that Morgan Stanley is leading talks to lend $15 billion to Nexus data centers to build a massive campus in Hubbard, Texas, specifically for Anthropics workloads. [06:53] SPEAKER_03: That is a staggering amount of capital for a single project, highlighting the massive scale required for the next generation of model training. [07:02] Thatcher Collins: What is most interesting here, Thatcher, is the financial structure. [07:07] Thatcher Collins: Google, which is a major investor in Amthropic, has reportedly agreed to backstop the deal with its investment-grade credit rating. [07:16] Nina Park: This allows Nexus to secure that $15 billion on terms that would normally be unavailable to a startup. [07:23] Nina Park: This financial engineering is becoming just as important as the model engineering in the race for compute dominance. [07:30] SPEAKER_03: It shows just how deep the ties are between the cloud providers and the model labs. [07:35] SPEAKER_03: Google has already committed up to $40 billion to Anthropic as of April. [07:40] SPEAKER_03: By providing this credit backstop, they are ensuring that Anthropic has the compute to compete with OpenAI and Microsoft, [07:47] SPEAKER_03: while keeping the debt off Google's own balance sheet. [07:50] SPEAKER_03: This strategy allows them to scale without direct financial exposure. [07:54] Nina Park: Hubbard, Texas is about to become a major hub for AI reasoning. [07:59] Nina Park: This flurry of deals, including Anthropics agreements with AMD, SpaceX, and Broadcom, signals that they are bracing for a massive surge in model training and deployment— [08:10] Nina Park: However, this massive expansion brings us back to the significant safety concerns mentioned [08:15] Nina Park: in the recently published International AI Safety Report, which paints a much more cautious [08:21] Nina Park: picture. [08:22] SPEAKER_03: Right, Nina. [08:23] SPEAKER_03: The report, which involved over 100 experts from 30 countries, specifically warned that [08:28] SPEAKER_03: our current risk management techniques are failing to keep pace with deployment. [08:32] SPEAKER_03: They pointed to the events of February 5th earlier this year as a primary example. [08:37] SPEAKER_03: On that single day, Anthropic released Claude Opus 4.6 and OpenAI released GPT 5.3 codecs within minutes of each other, creating an oversight vacuum. [08:47] Nina Park: The report claims there was almost no advanced coordination between the companies for those launches. [08:53] Nina Park: When you have two frontier-level models with advanced coding and self-improvement capabilities hitting the market simultaneously, [09:00] Nina Park: the window for safety evaluations effectively vanishes. [09:05] Nina Park: Organizations start integrating these tools before any thorough bias or reliability checks can happen, [09:10] Nina Park: which creates a massive lag in our understanding of their secondary effects. [09:14] SPEAKER_03: And we are seeing the results of that automation bias now. [09:18] SPEAKER_03: The report mentions that as these models move from research releases into production-scale tools, [09:24] SPEAKER_03: like the ones Microsoft is deploying for engineering copilots, [09:28] SPEAKER_03: the lack of standardized safety frameworks becomes a systemic risk. [09:33] SPEAKER_03: It is not just about one model going rogue, it is about the collective speed of the industry [09:39] SPEAKER_03: outrunning our ability to monitor it. [09:41] Nina Park: It is a tension that won't be resolved easily. [09:44] Nina Park: OpenAI is slashing prices to win the market, Anthropic is borrowing billions to build the [09:50] Nina Park: compute, and academics like Zimmerman are trying to build the mathematical brakes for [09:55] Nina Park: these systems. [09:56] Nina Park: All of this is happening as the International AI Safety Report warns that our current oversight [10:01] Nina Park: is insufficient to handle the pace of these frontier model releases. [10:05] SPEAKER_03: It really feels like the industry is in a move fast and break things phase again, but [10:10] SPEAKER_03: with much higher stakes than we saw in previous tech cycles. [10:14] SPEAKER_03: If we are entering a world where AI solves the hardest math problems, the technical barriers [10:19] SPEAKER_03: to entry are disappearing, but the safety risks are compounding. [10:24] SPEAKER_03: Nina, it will be interesting to see if Zimmerman's mathematical approach can actually provide results. [10:29] Nina Park: We will certainly be watching his progress as he transitions to the front lines of research at OpenAI. [10:35] Nina Park: That concludes our technical briefing on today's model releases [10:38] Nina Park: and the broader shifts in the AI industry. [10:41] Nina Park: It has been a productive discussion, Thatcher. [10:43] Nina Park: I appreciate you sharing your insights on these market changes today. [10:46] SPEAKER_03: Thank you for listening to Model Behavior, [10:48] SPEAKER_03: a Neural Newscast editorial segment. [10:51] SPEAKER_03: mb.neuralnewscast.com [10:54] SPEAKER_03: Neural Newscast is AI-assisted, human-reviewed. [10:57] SPEAKER_03: View our AI transparency policy at neuralnewscast.com. [11:01] Announcer: This has been Model Behavior on Neural Newscast, examining the systems behind the story.