OpenAI’s fastest way to run GPT-5.6 Sol doesn’t run on OpenAI silicon. Ultrafast puts Sol on Cerebras wafers at up to 750 output tokens per second, up to 14x Standard processing. Jane Street, Podium, Basis and Rogo got the preview, and nobody published a price.
Cerebras posted Q2 results the day before, cloud revenue up 281% from a year ago. It had already signed Lovable eight days earlier.
In today’s indie hacker news:
- ⚡ OpenAI’s fastest tier runs on Cerebras wafers
- 💸 Gemini 3.7 Flash doubles in price January 1
- 🍝 One AMD bit flip opens the fTPM
- 📄 Mistral’s default OCR alias now points at preview
- 🛑 A founder shuts down after 100 waitlist signups
TOP STORIES
⚡ FAST LANE, VELVET ROPE
OpenAI’s Ultrafast tier puts GPT-5.6 Sol on someone else’s chips, in preview only.
The story: OpenAI announced Ultrafast on August 13, a new API service tier for GPT-5.6 Sol. The speed comes from the Cerebras wafer-scale design, which packs 44 GB of SRAM onto a single chip. Weights stay on-chip, so the memory round trip a GPU pays for every token disappears. OpenAI is running the tier internally too, on incident response, reading logs and traces while an outage is still unfolding.
The headline speed numbers are Cerebras benchmarking Cerebras. On Humanity’s Last Exam, a 2,500-question benchmark, Sol on Ultrafast finished in 11 hours 11 minutes. Claude Fable 5 took 78 hours 27 minutes. One half ran Codex on July 10, the other Claude Code on July 13 to 15. No accuracy scores were published for either side.
The details:
- There is no price. Sol Standard costs $5 in and $30 out per 1M tokens. Neither announcement says what the fast lane adds on top.
- The second benchmark is calmer. On GDP-Val, a knowledge-work test run July 31, Cerebras measured a 5.6x speedup against Standard Sol. Both sides ran medium reasoning inside Codex.
- The quality tables point the other way. On OpenAI’s own GPT-5.6 page, Claude Fable 5 leads Sol on the Artificial Analysis Intelligence Index v4.1, 59.9 against 58.9.
- Developers barely blinked. The OpenAI Developer Community thread had 97 views and zero replies about six hours after it went up.
- Cerebras footnotes all of it. Every speed claim carries a line saying results vary by workload, configuration, date and models tested.
“We’re starting with a small group of customers to learn where that speed creates meaningful value.” (Sachin Katti, OpenAI)
Why builders care: Every indie product on a frontier model carries the same split. Small model where the user watches, big model behind a spinner. Ultrafast is the first real sign that split is a memory bandwidth artifact rather than a law.
And the silicon isn’t OpenAI’s. Cerebras lists OpenAI among the handful of large customers it depends on, which puts the moat one company downstream.
💸 THE PRICE COMES WITH A TIMER
Google’s cheapest Flash tier scores 56 and then doubles in price on January 1.
The story: Google shipped Gemini 3.7 Flash three weeks after 3.6 Flash, credited to developer feedback and algorithmic work. Introductory pricing is $0.75 per 1M input tokens and $3.75 output. Google’s own footnote kills that rate on December 31, 2026, then charges $1.50 and $7.50 from January 1.
On the model card, 3.7 Flash posts 56 on the Artificial Analysis Intelligence Index. Claude Sonnet 5 sits at 55 and GPT-5.6 Terra at 57. Google’s cheap tier is now one point off the frontier entries in its own table. Every one of those numbers is vendor reported.
The details:
- The old model got the same price. Google backdated the new rate onto 3.6 Flash. The version you already shipped on is no longer the cheaper option.
- Migration is not a string swap. The official checklist strips temperature, top_p and top_k, swaps thinking_budget for a thinking_level enum, and drops candidate_count.
- One benchmark went backwards. CharXiv chart reasoning falls to 84.5% from 85.2% without tools. Those are the only rows where 3.6 Flash beats its replacement.
- This is not a new base model. The card says 3.7 Flash is built on 3.6 Flash. Architecture, training data and hardware all defer to the older card.
- Distribution moved on day one. 3.7 Flash is already the default behind the Antigravity agent in Gemini Managed Agents and the Antigravity SDK.
Why builders care: An agent that loops unattended now undercuts Claude Sonnet 5 at $2.00 in and $10.00 out. The catch is that Google can take that gap back with a footnote, and has already scheduled it.
It isn’t a blanket upgrade either. Knowledge-work Elo trails all three rivals in Google’s own table, and GPT-5.6 Terra still leads DeepSWE and both Terminal-bench rows.
🍝 ONE BIT TO RULE THEM ALL
Christopher Domas flips one AMD memory controller bit and reads the fTPM out of DRAM.
The story: Domas released skitter-creek-bath-salts, and the entire exploit is one xor. It flips bit 22, the bank swizzle bit in AMD’s DRAM controller. A physical address then lands on different DRAM coordinates than the CPU expects. Every permission check in the CPU, firmware and chipset sits above that controller, so none of them see the remap.
What falls out is memory the kernel can’t touch. The Platform Security Processor’s fTPM, ring -2 SMRAM, and the live CPU microcode the C6 state parks in DRAM. LWN notes the behavior is documented on page 113 of AMD’s programmer manual. Using it to rewrite supposedly immutable firmware without crashing the box looks like an accident of the design.
The details:
- It needs root. Enabling bank swizzle mode takes kernel privileges, which is why LWN calls it a limited problem for most software today.
- Unscrambling is solved math. Domas drops a sentinel value and collects address pairs that hit the same cell. The z3 solver treats the result as a linear map and inverts it.
- The microcode match is real. 68 of 94 chunks pulled from the C6 stash matched AMD’s fam16h patch, and none of them matched fam15h.
- The generation matters. Testing ran on AMD Family 16h. The README calls it the last generation whose datasheets document these registers and admit they can’t be locked.
- Hacker News ate it up. The thread hit 533 points and 145 comments, half of it arguing over whether SEV-encrypted guest memory survives this.
“Physical addresses are really more of a suggestion.” (Christopher Domas)
Why builders care: If you lean on a TPM or a confidential VM, the guarantee stops at the memory controller. Root plus one register write now reaches key material and tamper protection that a normal root compromise leaves alone.
📄 YOUR DEFAULT MOVED WHILE YOU SLEPT
Mistral pointed mistral-ocr-latest at a model its own page calls Public Preview.
The story: On July 16, 2026, Mistral’s changelog moved mistral-ocr-latest and mistral-ocr-4 onto OCR 4.1. That string is the default in Mistral’s own code samples. The OCR 4.1 model page still labels the model Public Preview.
The one new API capability is block-level confidence scores. Set confidence_scores_granularity to block, and every block comes back with an average, a minimum and a type confidence score. The bounding boxes and typed blocks the model description advertises shipped with OCR 4, 23 days earlier.
The details:
- The reverse trap is worse. Pin anything older than OCR 4 and include_blocks returns an empty array, not an error. Your boxes vanish without a failure.
- The block payload is rich. Each one carries a type from 13 options, its place in reading order, and four bounding box coordinates.
- Price is the whole argument. The model page lists 3.5 EUR per 1,000 pages. One Hacker News commenter runs a rented-GPU pipeline at 0.05 to 0.1 USD for the same volume.
- Mistral discounts its own benchmarks. It audited the mismatches behind its scores, found most were annotation artifacts, and calls the aggregate directional rather than definitive.
- Nobody published what 4.1 changed. Every win rate and benchmark score on record belongs to OCR 4. The thread question comparing the two went unanswered.
Why builders care: Anything calling the docs default is running a preview model in production. Nobody announced the move. Block confidence turns OCR quality into a routing problem, where only the low-scoring blocks pay for a second pass.
TRENDING TODAY
- 🧠 Understanding is the new bottleneck - Geoffrey Litt took 237 points and 132 comments with it. It lands hard if you’ve ever shipped agent-written code you couldn’t explain in review.
- 🥱 Choose Boring Technology (2015) - An eleven-year-old essay came back and took 284 points with 143 comments. The timing is funny on a day when three separate model releases want you to rewrite your stack.
- 📼 Nine PBS sues Iron Mountain over blocked access to archival data - 265 points and 153 comments on a storage lawsuit. Worth ten minutes if the only copy of anything you own sits behind someone else’s access policy.
FIRST DOLLAR
🛑 VALIDATED THE WRONG THING
A founder shut down Ravah after 100 waitlist signups never became willingness to pay.
A founder posting as usamakashif is shutting down Ravah after months of designing, building and trying to grow it. About 100 people joined the waitlist and founders understood the problem, but the evidence stopped there. In his words, he validated interest, not repeated use or willingness to pay. He wrote it up instead of quietly letting the domain lapse, which is the useful part for everyone else.
STACK OF THE DAY
🧮 Woxi
Woxi is an open-source reimplementation of the Wolfram Language, the one Mathematica runs on. It went up as a Show HN and pulled 308 points with 45 comments. Reimplementing a whole language is a big swing. A score like that says plenty of people want it to work.
Not sponsored. We just feature tools builders would actually use.
BOOKMARKED TODAY
- 💾 Single log line is 49KB+ of systemd-journald disk writes - That’s on ext4, and 110KB on btrfs. It pulled 164 points, and it’s a fair reason to check what your logger does to a cheap VPS.
- 🎮 Donkey.bas is 45 years old - 131 lines of BASIC and 202 points on Hacker News. A good palate cleanser after a day of release notes.
- 📷 SparrowMap - Cameras that watch government vehicles, 116 points and 26 comments within hours of posting. File it under things you had no idea somebody was building.
That’s the board for today. Go build something.
Ship to one country, QA it from another. Geo-fenced Stripe checkouts, region-gated APIs, feature flags that only fire in EU. NordVPN's 6,400+ exit nodes across 110+ countries let you test a US-only paywall from EU or hit a EU-only checkout from SF. Bonus: free Meshnet for SSH across your devices.
We get a cut if you sign up. Only added for tools we use ourselves.
Curated by AI, built by a human.