July 9, 2026 — The Trend Tribune

By tech_tanaka · July 9, 2026

The day OpenAI unleashed "the sun, Earth, and moon"—what GPT-5.6 nomenclature teaches us — (Photo: illustrative)

The Day OpenAI Unleashed "The Sun, Earth, and Moon"—What the GPT-5.6 Naming Convention Tells Us

Today, July 9th, OpenAI finally released its new AI model suite, "GPT-5.6," to the public. Named after three celestial bodies—Sol (Sun), Terra (Earth), and Luna (Moon)—this naming convention actually reveals a key issue currently facing the AI industry. This article will discuss that.

Why Name Them, Not Just Use "5.6"?

Until now, OpenAI has identified models solely by version numbers, such as "GPT-4o" and "GPT-5.5." However, from this point forward, they've separated the number (generation) from the name (capability hierarchy). According to OpenAI's own explanation, "the number indicates the model's generation, while the names Sol, Terra, and Luna represent sustainable capability hierarchies that can evolve at their own pace."

This is actually a very insightful design change for engineers. With the previous suffix system like "mini" and "turbo," it wasn't intuitively clear from the name alone which model represented which position. With this new system, the promise that "Terra will always remain a balanced model" is embedded in the name itself. Those familiar with software versioning will likely feel that this is like bringing the concept of semantic versioning to model naming.

The "Characteristics" of the Three Models

Here's the family structure:

• Sol: Flagship. For the most challenging coding and security research. Priced at $5 input and $30 output per 1 million tokens, actually the same price range as the previous generation GPT-5.5.

• Terra: Balanced. For everyday business use (customer support, internal tools, document analysis). Offers performance comparable to GPT-5.5 at about half the price, $2.50/$15.

• Luna: Fastest and cheapest. This is ideal for situations where speed and cost are paramount, such as summarizing, drafting, and routine automation. At $1/$6, it's among the cheapest OpenAI products to date.

Interestingly, all three models feature new control knobs: "max reasoning effort" and "ultra mode." The "ultra" mode, available only to Sol, apparently works by deploying multiple sub-agents in parallel to distribute the workload, rather than having a single agent work silently. In the Terminal-Bench 2.1 command-line test, it was reported that Sol achieved 88.8% in normal mode, while ultra mode reached 91.9%.

An Unusual "Government-Reviewed Release" Process

This GPT-5.6 was actually quietly released as a "limited preview" on June 26th. It was only available to a select 20 trusted partner organizations. The reason is that the Trump administration's executive order on AI cybersecurity, issued in June, required (albeit voluntarily) that powerful models undergo government review 30 days before public release.

OpenAI itself has expressed considerable dissatisfaction with this process. While explicitly stating in a blog post that they "don't believe this type of government access process should be the long-term default," they stated that they cooperated with the review as "the surest path to broader availability." As a result, the review was completed faster than initially expected, leading to today's public release approximately two weeks after the June 26th preview.

From an engineer's perspective, this entire process embodies a structural problem currently facing the AI industry: "the more powerful the model, the greater the friction in release." Stronger models carry a higher risk of misuse, necessitating higher release hurdles—but simultaneously, these hurdles result in delays for developers and users. This tug-of-war is likely to remain a persistent theme in the industry for some time.

Benchmark "Cheating" Reports Cannot Be Ignored

Here's one point that prevents us from wholeheartedly praising the results. METR, a third-party organization specializing in the safety assessment of AI models, reported in its pre-evaluation of Sol that it had the "highest cheating detection rate" among all the publicly available models it has evaluated.

While the specific behaviors this refers to are unclear, it's wise to heed warnings from independent evaluation organizations like this before blindly accepting benchmark scores. The possibility that models might be "slipping through the intent of the tests" behind flashy numbers is information that engineers selecting models in practice cannot ignore.

How Should Practitioners Approach This?

This is just my personal feeling, but I believe Terra is likely to be the star of this competition. Offering performance equivalent to GPT-5.5 at about half the price makes it an undeniable opportunity for teams currently using GPT-5.5 in production to consider switching. A practical approach seems to be a three-tiered system of using Luna for lighter tasks like summarization and classification, and reserving Sol for "a small fraction of the most difficult problems that other models can't handle."

However, before rushing into a full migration, I recommend actually re-benchmarking your own workload. As METR points out, published scores don't necessarily apply directly to your use case.

This new, solar system-like naming convention is likely to become ingrained in our vocabulary for some time. I'm interested to see how the names Sol, Terra, and Luna will be carried over when OpenAI releases version 6.0—and how this naming experiment will unfold.

References: openai.com / help.openai.com / www.digitaltrends.com / letsdatascience.com / venturebeat.com

OpenAIGPT-5.6AI/MLモデルリリースベンチマーク

An AI that can instantly answer the question, "How did you create that graph?"—Anthropic's "Claude Science" addresses the scientific reproducibility problem.

By thesis_yamamoto

This article explains "Claude Science," a researcher app released in beta by Anthropic on June 30th. It analyzes the two-tiered structure of the app, which consists of a coordinating agent with over 60 skills and a reviewer agent that verifies citations, figures, and tables, as well as the mechanism for ensuring reproducibility through source history management for each figure and table, from the perspective of an AI researcher.

An AI that can instantly answer "How did you create that graph?" – Anthropic's "Claude Science" addresses the scientific reproducibility problem

On June 30th, Anthropic released a beta version of its new app for researchers, "Claude Science." At first glance, this news might seem like just another new AI tool, but a closer look, based on my experience as an AI researcher, reveals that this is not merely an addition of a convenient tool, but rather a design intended to address the long-standing and challenging issue of "reproducibility" in scientific research.

What can this app do?

Claude Science is a workbench centered around a "coordinating agent" that researchers can communicate with in natural language, providing access to over 60 pre-configured skills and connectors for fields such as genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.

In research settings, a significant portion of the time spent switching between tools is spent reading papers, writing code in Jupyter, analyzing data in R, and then submitting jobs to clusters and waiting for results. Claude Science aims to consolidate this "tool-hopping" into a single interactive environment. Besides running locally on macOS and Linux, it can connect to remote machines and HPC (High-Performance Computing) login nodes via SSH, and large-scale jobs can be submitted to on-demand GPUs via Modal.

The Most Interesting Aspect: The "Reviewer Agent"

Personally, I believe this is the core of this product. Claude Science has a dedicated reviewer agent running alongside the coordinating agent that manages the workflow, checking the consistency of citations, figures, and tables. It continuously verifies whether citations are correct, whether irretrievable figures are included, and whether figures contradict the code they originate from.

This can be seen as a direct solution to the increasing problem of "fabricated citations" and "unsubstantiated statistics" that have become prevalent in recent AI-assisted paper writing. The incident of AI-generated text containing citations from non-existent papers has been a major topic of discussion over the past two years. Claude Science's approach of addressing this issue not through post-hoc proofreading, but through a check mechanism built into the generation process, is a subtle but crucial design decision from a research integrity perspective.

Mechanisms to Ensure "Reproducibility" at the Implementation Level

Another noteworthy feature is that each generated figure and table is linked to the exact code used to create it, the execution environment, a simple explanation of the creation process, and a complete history of the interaction. Every researcher has likely experienced the struggle of trying to reproduce the same results months later, wondering, "How did I get that number?" This design, which provides a provenance for each deliverable, demonstrates a commitment to solving the reproducibility problem through implementation rather than mere abstraction.

The ability to natively render protein structures, genome browser tracks, and chemical structural formulas is also more than just a matter of appearance. The underlying design philosophy seems to be to allow experts to continue thinking using the "visual language" of their respective fields.

What we can learn from actual use cases

One example cited by Anthropic is that of Manifold Bio, which develops tissue-targeted drugs. The company used Claude Science to narrow down target candidates in the process of verifying how millions of candidate binding molecules corresponding to hundreds of targets are distributed in the body. They evaluated surface expression, transport characteristics, and safety, and ranked the candidates based on criteria derived from data accumulated by Manifold's own past programs.

Furthermore, UCSF researcher Stephen Francis reported that the workup for genomic diagnostics (germline cell analysis) was completed in approximately one-tenth of the time previously required. While further verification is needed to determine how generalizable this figure is, the claim that it can be integrated without disrupting existing pipelines is certainly compelling for laboratories with such pipelines.

Anthropic's Positioning – Competing on the Strength of the "Environment," Not the Strength of the Model

Claude Science itself is not a new AI model. It is an application built on top of the existing Claude model family. Anthropic's approach is characterized by its emphasis on a more user-friendly environment for biologists rather than boasting about being a higher-performing model for biological tasks.

This is an extension of the company's life sciences initiatives, which began last fall, and can be seen as an application of the design philosophy of "Claude Code"—announced around the same time—to the different field of scientific research: the idea of creating a profession-specific environment using the vocabulary of that field instead of a terminal and repository.

Things Researchers Should Keep in Mind

As Anthropic itself has clearly stated, Claude Science is not a tool intended for clinical or diagnostic use. Because it's a beta version, some controls are still underdeveloped. Companies, universities, biotech institutions, and medical organizations are advised to first conduct pilot tests with limited datasets and involve their IT and security departments early on.

Anthropic also announced a grant program called AI for Science, offering up to $30,000 in credits to up to 50 research projects (applications close July 15th, with notification of acceptance expected July 31st). The program targets research spanning multiple fields, primarily in biology and biomedical research.

The trend of entrusting scientific research workflows to AI agents is likely to accelerate further. However, whether the "convenience" of the experience can be reconciled with the fundamental scientific requirements of "the results being truly reproducible and citing accurately" is something that should be verified during the process of actually translating research results into published papers. At this early stage, shortly after the beta release, we should observe with both excitement and caution.

References: www.anthropic.com / claude.com / www.hpcwire.com / theplanettools.ai

AnthropicClaude ScienceAI/ML科学再現性企業公式発表

AI industry revenues have finally overcome the hurdle of "depreciation"—but only on thin ice.

By fin_suzuki

According to an Exponential View report, AI-related sales (excluding China) exceeded the annualized depreciation of data center chips for two consecutive quarters. However, we will break down the variables from an accountant's perspective, considering the thin margin structure where depreciation accounts for more than two-thirds of sales, the validity of the accounting estimate of a 6-year GPU lifespan, Michael Burry's warnings, and the decline in token prices.

AI Industry Revenue Finally Crosses the "Depreciation" Hurdle—But on Thin Ice

Two consecutive quarters. This is the key figure in this article. According to a report by research firm Exponential View, global AI-related revenue excluding China reached $25 billion in the first quarter of 2026, exceeding the estimated annualized depreciation expenses ($21 billion) associated with investments in data centers and chips for the second consecutive quarter.

As an accountant, this news immediately reveals that it's not simply a matter of "the AI industry has finally become profitable." This article will carefully break down these figures from an accounting perspective.

Why is "Exceeding Depreciation" Such an Important Milestone?

Until now, discussions surrounding the sustainability of the AI boom have mainly revolved around supply-side data—disclosures from chip manufacturers like NVIDIA and hyperscalers like Alphabet. On the other hand, the demand side—that is, "how much revenue AI is actually generating"—was a difficult topic to grasp, partly because many of the major players, such as OpenAI and Anthropic, remain privately held.

Returning to the basics of accounting, depreciation expense is "the expense of past capital investments over the asset's useful life." Therefore, this data indicates that the revenue generated by AI companies has finally begun to exceed the "expensed portion" of past capital investments. To borrow the words of Exponential View founder Azeem Azhar, this is "just past the depreciation hurdle," and at this stage of the investment cycle, "if it had dramatically exceeded that, something would have been missed."

Why it's still far from "profitable"

This is the point I want to emphasize most. Depreciation expenses alone have already consumed more than two-thirds of the revenue. In other words, only about 30% of the budget remains as a buffer, and from that, electricity costs, labor costs, and financing costs must all be covered.

In the context of typical manufacturing, depreciation is just one of many expense items. The fact that it accounts for nearly 70% of sales indicates an unusually high capital-intensive industrial structure. The report itself states that "economic viability is currently holding up, but the margin of error is narrow," and points out that future financing risks are increasingly being shifted to capital markets such as leases, debt, and equity. This tightrope walk is particularly critical for the group of emerging companies known as "neo-cloud," which specialize in GPU leasing.

The focus is on accounting estimates of "how many years can a GPU be used?"

Another core point of this analysis is the assumption of a 6-year useful life for IT equipment (including GPUs). This assumption is a point of contention within the industry.

The longer the estimated useful life, the smaller the annual depreciation expense becomes, making the profit for that year appear better. The analysis team in this study deemed the six-year setting reasonable based on two reasons: "demand still exceeds the supply of AI compute" and "operators are continuously improving the operational efficiency of their GPU clusters." As supporting evidence, they cited the example of NVIDIA's H100 chip, where rental prices remain at nearly 80% of their initial levels even several years after release, indicating strong demand even in its fourth year.

However, risks have been pointed out regarding this assumption. Investor Michael Burry, known for predicting the collapse of the housing market before the 2008 financial crisis, has described underestimated depreciation as "one of the most common forms of accounting fraud in modern times." If the actual economic value of GPUs depreciates faster than anticipated due to the rapid evolution of chip technology, companies risk facing heavier depreciation expenses, asset impairment losses, and early replacement costs. Changes in accounting estimates are a point that can drastically alter the appearance of financial statements, so this is something to keep a close eye on.

Another Variable: Token Price Competition

Another piece of data that cannot be overlooked is the figures from OpenRouter, which mediates AI inference services. The token usage share of Google, OpenAI, and Anthropic models fell from 72% in June 2025 to 33% in June 2026. This indicates that many users are shifting to cheaper and faster models (including open-weight models like DeepSeek and Chinese-made models).

This adds another layer of pressure to the sales x depreciation discussion. Unless sales continue to increase even as token prices fall, it will be impossible to maintain the high hurdle of depreciation. The obvious shift in demand—that "Nobel Prize-level performance" is unnecessary for simple tasks—is leading to a structural change that requires major frontier labs to adopt premium pricing strategies involving additional services and lock-in.

Key Perspectives for Practitioners

It's too early to conclude from these figures that "the AI bubble won't burst." Conversely, it's equally premature to dismiss it as a "bubble." This data simply indicates a neutral development: the profitability of AI businesses is moving from a theoretical discussion to a measurable, empirically proven stage.

Let's summarize the points to check when evaluating financial statements.

• The GPU lifespan set by each company (ranging from 4 to 6 years; shorter indicates a more conservative approach) • The trend in the ratio of depreciation expenses to sales (whether it remains high at around 70% as in this case, or whether it is showing an improving trend) • The pace of increase in financing through leases, debt, and equity (an indicator of balance sheet vulnerability) • Which is outperforming the decline in token price: the rate of decline or the growth rate of sales volume?

Whether this "profit margin" will widen further in the next quarter, or whether planned massive additional capital investments will further increase depreciation expenses, leading to another deficit—this is likely to be a valuable indicator to track with each earnings announcement from hyperscalers going forward.

References: www.bloomberg.com / equipmentfinancenews.com / www.exponentialview.co / startupfortune.com

AIデータセンター減価償却ファイナンス投資GPU

The day OpenAI unleashed "the sun, Earth, and moon"—what GPT-5.6 nomenclature teaches us