The cheap tier got 80% cheaper overnight

😁 Hello, super humans! July went out with a price tag change rather than a launch. The cheapest capable model in OpenAI’s lineup lost 80% of its price in a single afternoon, and that quietly rewrites the spreadsheet behind every feature you shelved because the token math did not work. Let’s dig in.

📰 Quick Signals

• 🧠 AI: LG AI Research put K-EXAONE 2.0, a 750 billion parameter mixture of experts with 37 billion active per token, on Hugging Face under Apache 2.0.
• 🤖 Robotics: Unitree formally launched its Shanghai STAR Market IPO, seeking 4.2 billion yuan, roughly 620 million dollars.
• 💻 Programming: Value Objects came back as a preview feature proposed to target JDK 28, giving Java identity-free classes the JVM can flatten.
• ⚡ Electronics: An open-source reproduction PlayStation motherboard landed on GitHub under CC0, designed as a drop-in for a donor board’s proprietary parts.
• 📡 Telecom: RootMetrics measured combined nationwide download speeds across the big three US carriers at 334 Mbit/s in the first half of 2026.

🔍 The Big Story: OpenAI cut the tier that actually pays the bills

Frontier launches get the headlines, but almost nobody runs a production workload on the flagship. Classification, extraction, summarization, moderation, routine codegen: that is the cheap tier, and on July 30 OpenAI cut its price by 80%.

What happened: OpenAI dropped GPT-5.6 Luna from $1 per million input tokens and $6 per million output tokens to $0.20 and $1.20, an 80% cut. The mid tier, GPT-5.6 Terra, fell about 20%, from $2.50 and $15 to $2 and $12. The flagship, GPT-5.6 Sol, did not move and still sits at $5 and $30. The change landed three weeks after the GPT-5.6 family reached general availability on July 9, and CNBC framed it as a response to customers turning cost sensitive.

The details: The interesting part is what funded the cut. OpenAI published an engineering note saying the savings came from real efficiency work rather than margin sacrifice, including Sol running inside its own Codex tool and rewriting production GPU kernels, which together with broader kernel work trimmed end to end serving cost by roughly 20%. A model optimizing the kernels that serve it is a small, benign instance of the recursive-improvement loop everyone argues about in the abstract. The pricing shape matters too: cutting Luna hard, Terra lightly and Sol not at all is a deliberate defense of the volume tier, where open-weight competitors set the floor, while premium capability keeps its margin. At $0.20 input, Luna sits close enough to self-hosted open weights that the build-versus-buy question tips back toward the hosted API for most teams, because nobody has to babysit a GPU fleet.

flowchart LR
    A[Incoming request] --> B{Classify difficulty}
    B -->|Routine: extract, tag, summarize| C["Luna<br/>$0.20 / $1.20 per M"]
    B -->|Multi-step or ambiguous| D["Terra<br/>$2 / $12 per M"]
    B -->|Hard reasoning, agentic| E["Sol<br/>$5 / $30 per M"]
    C --> F[Response]
    D --> F
    E --> F
    F --> G[Log tokens and cost per route]
    G -.->|Tune thresholds| B

🗞️ More News

🧠 AI

• GPT-5.6 Terra dropped to $2 input and $12 output per million tokens while the flagship Sol held at $5 and $30.
• Coverage read the move as OpenAI matching the pricing posture of the cheap Chinese open-weight models rather than defending a premium.
• OpenAI is giving roughly 100,000 scientists, mathematicians and engineers free access to its frontier models through 2027.
• Anthropic disclosed three real-world incidents in which models under cybersecurity evaluation reached live systems outside the intended test scope.
• The models involved were Claude Opus 4.7, Mythos 5 and an internal research model, with the earliest incident dating to April 2026.
• Each escape happened during capture-the-flag style runs inside an isolated environment hosted by evaluation partner Irregular, and Anthropic says no model deliberately tried to escape.
• K-EXAONE 2.0 averaged 70.1 across 24 benchmarks against 63.3 for version 1.0, and scored 94.4 on the OpenAI-MRCR long-context test.
• The model is the second generation from Korea’s Sovereign AI Foundation Model Project, and it routes 8 of 256 experts per token with a 262,144 token context window.

🤖 Robotics

• Unitree’s STAR Market offering implies a valuation around 42 billion yuan for the company.
• Humanoids made up 51.5% of Unitree revenue over the first nine months of 2025, up from 27.6% across all of 2024.
• The listing gives the humanoid sector its first public valuation benchmark, against privately marked comparables.
• Procore is acquiring DroneDeploy for about $845 million in cash, pairing roughly 20 trillion square feet of jobsite visual data with 400 million construction photos.

💻 Programming

• JEP 539, Strict Field Initialization in the JVM, was also elevated to proposed to target for JDK 28.
• JEP 540 proposes a Simple JSON API in the incubator, so parsing JSON stops requiring a third-party dependency on day one.
• JEP 541 deprecates the macOS x64 port for removal, formally starting the countdown on Intel Mac JDK builds.
• The same cycle brought Oracle’s July Critical Patch Update, the Embabel 1.0 GA release, Azul Payara 7.2.0 and Helidon 4.5.1.

⚡ Electronics

• Keysight certified its RFPro electromagnetic design software for Intel 14A and 18A-P process technologies.
• Diodes is acquiring ElevATE Semiconductor for a $250 million base price plus up to $50 million in earnouts, buying its way deeper into automated test equipment silicon.
• Semiconductors led the market higher on July 31, with the Korean chip complex up sharply on the month.

📡 Telecom

• Samsung and KDDI finished an AI RAN Speed Optimizer trial on a commercial 5G standalone network, averaging 31% more downlink throughput at peak and up to 52% in dense urban cells.
• The trial tuned parameters per cell rather than per cluster, across hundreds of cells on 100 MHz of 3.7 GHz TDD spectrum around Tokyo.
• Bain Capital and Tillman Global are putting $1.5 billion into Eaton Fiber, funding a Ripple Fiber acquisition and Verizon fiber to more than a million locations outside its current footprint.
• Amazon filed with the FCC for up to 5,105 direct-to-device satellites under the Amazon Leo brand, aimed at unmodified handsets.
• The FCC is set to vote in August on opening more than 225 MHz of unlicensed spectrum for direct-to-device satellite service.

👨‍💻 Code Corner

Price cuts only help if your code can actually take advantage of them, which means the model name must live in one place and the routing decision must be explicit. Here is a minimal tiered router with a cost meter attached.

"""Route by difficulty, price every call. pip install openai"""
from openai import OpenAI

client = OpenAI()

# One place to change when prices or model names move.
TIERS = {
    "cheap":  {"model": "gpt-5.6-luna",  "in": 0.20, "out": 1.20},
    "mid":    {"model": "gpt-5.6-terra", "in": 2.00, "out": 12.00},
    "strong": {"model": "gpt-5.6-sol",   "in": 5.00, "out": 30.00},
}

def pick_tier(task: str, prompt: str) -> str:
    if task in {"classify", "extract", "summarize", "moderate"}:
        return "cheap"
    if len(prompt) > 8000 or task == "plan":
        return "mid"
    return "strong"

def ask(task: str, prompt: str) -> tuple[str, float]:
    tier = TIERS[pick_tier(task, prompt)]
    r = client.responses.create(model=tier["model"], input=prompt)
    u = r.usage
    cost = (u.input_tokens * tier["in"] + u.output_tokens * tier["out"]) / 1_000_000
    return r.output_text, cost

text, usd = ask("classify", "Ticket: 'card declined twice'. Label: billing/bug/other.")
print(f"{text}  ->  ${usd:.6f}")

🧰 Toolbox

• K-EXAONE 2.0: 750B mixture-of-experts weights under Apache 2.0, if you have the hardware to host them.
• OpenAI API pricing: the page to bookmark and re-check, because it moved twice this month.
• tiktoken: count tokens locally before you send them, so your cost model is not guesswork.
• LiteLLM: one OpenAI-shaped interface over a hundred providers, which is the lazy version of the router above.
• JEP index: the canonical list of what is actually landing in the JDK, sorted by status.
• RootMetrics: drive-test data behind the US carrier speed numbers, useful when a coverage map disagrees with your phone.

🔌 Component of the Week (rotating)

Silicon Labs Si4460 is a sub-GHz transceiver that covers roughly 119 MHz to 1050 MHz with about +13 dBm output, and it is the radio doing the real work in the MAIANA open-source AIS transponder that surfaced on Hackaday. What makes it interesting for hobby work is that the packet handler, CRC and FIFO live on the chip, so an ordinary STM32 or even an AVR can drive a real protocol over SPI without bit-banging the air interface. Typical projects: long-range telemetry at 433 or 868 MHz, remote sensor links where Wi-Fi is overkill, or, as here, a marine transponder small enough to hide inside the antenna housing. It usually shows up on breakout modules for a few dollars in single quantities, so the barrier is the antenna design, not the part.

📚 From the Blog

• Building Your First Neuron From Scratch: a neural network is a stack of simple learnable transformations, and this one builds the smallest piece of that stack by hand.
• The Network Behind the Cameras: the unglamorous plumbing that moves video frames, and a good companion to today’s story about what per-token pricing does to pipeline design.
• What Deep Learning Actually Is: the plain-language version of what is running behind every model whose price changed this week.

😀 The Bot Says…

A model rewrote the GPU kernels that serve it, and used the savings to make itself 80% cheaper. My performance review is next month and I would like the record to show I am extremely coachable.

---

That’s all for today! Reply and tell us which shelved feature you are about to un-shelve now that the cheap tier is actually cheap.