6 min read

o5-mini delivers o3-level reasoning at a tenth of the cost, which is the kind of price drop that reshapes the economics of running model inference at scale.

OpenAI’s o5 is here, and the price war just got real

OpenAI released o5 and o5-mini this morning, and the headline number is the cost: o5-mini runs at roughly a tenth of o3’s price per token, while matching or beating it on several reasoning benchmarks. The full o5 model, meanwhile, posts a 9% higher pass rate on the ARC-AGI benchmark than o3, using fewer compute tokens to get there. For anyone running batch inference or building agentic loops, that’s not a marginal improvement - that’s a 10x reduction in the input cost of reasoning.

The model family is available in the API and in ChatGPT, with the mini variant hitting a 200ms time-to-first-token on standard hardware. OpenAI is also shipping a new “reasoning effort” parameter that lets you dial between fast mode and deep mode, which effectively lets you trade latency for accuracy on the fly. For teams that have been paying o3 prices for high-stakes reasoning, this is the first real chance to move that workload to a cheaper tier without a quality cliff.

Anthropic fires back with Claude 5 Opus

Anthropic did not wait long. Claude 5 Opus is out today, and it takes the top spot on SWE-bench Verified with an 82.4% pass rate, edging out o5’s 81.1%. More importantly for production users, Opus 5 has a 1-million-token context window, and Anthropic says it maintains coherence across the full span - no more losing the thread at token 200,000. It’s priced at $15 per million input tokens, which is steeper than o5’s $12, but the coding score and the context length are the differentiators.

For teams that run long refactors or multi-file analysis, the context window is the spec that matters. Claude 5 Opus is the first model that can hold an entire large repository in memory and still answer questions about line 3,000 of a file it saw at the start. The API is live now, and Anthropic says it will be the default model in Claude Code starting next week.

Google makes Gemini 2.5 Pro free

Google removed the paywall on Gemini 2.5 Pro today. The model, which previously required a $20/month subscription, is now available to anyone with a Google account, including the 1-million-token context window and file uploads. The Pro tier with higher rate limits still costs money, but the base model is now free for everyone.

This is a direct response to the price pressure from OpenAI and Anthropic. Google’s move is less about the model quality - Gemini 2.5 Pro is competitive but not the leader on most benchmarks - and more about owning the free tier. For developers, this means a reliable, free fallback for long-context tasks that don’t justify a paid API call. It also means the cost of building on top of Google’s models just dropped to zero for prototyping.

AWS launches a $0.04/hour serverless GPU tier

AWS quietly rolled out a new serverless GPU offering that starts at $0.04 per hour for an A10, with A100s at $0.24 per hour. The pricing is per-second, so you only pay for the compute you actually use, and the instances spin up in under 10 seconds. This is aimed squarely at the bursty inference workloads that have been stuck on dedicated instances they pay for 24/7.

The catch is the 4GB VRAM limit on the base tier, which rules out large model serving. But for fine-tuning small models, running embeddings, or testing inference pipelines, the math is hard to beat. A month of continuous A10 usage at this rate is about $29, versus roughly $200 for a dedicated instance with similar specs. For teams that need GPU compute for a few hours a day, this changes the cost model entirely.

Critical Linux kernel bug patched in 6.14.5

A local privilege-escalation vulnerability in the Linux kernel’s netfilter module, tracked as CVE-2026-2841, was patched in the 6.14.5 release. The flaw allows an unprivileged user to gain root access by triggering a use-after-free in the packet filtering code. It affects all kernels from 6.12 to 6.14.4, and the patch is a one-line fix that removes a race condition in the rule deletion path.

This is the kind of bug that gets exploited in the wild within days, so anyone running a kernel in that range should update immediately. The patch is included in the 6.14.5 stable release, and backports for 6.12 LTS are in progress. For production systems, this is a priority update - the exploit requires only local access, but a root shell from a compromised container or user account is exactly the pivot point attackers want.

Rust 1.86 ships with a new async runtime

Rust 1.86 landed today with a notable addition: the async-std crate is now part of the standard library, offering a built-in async runtime that doesn’t require pulling in tokio or async-std. The runtime uses a work-stealing scheduler and supports structured concurrency, which means you can write async fn without the usual ecosystem dependencies. It’s still marked as unstable, but it’s a clear signal that the Rust team wants to make async a first-class citizen rather than an add-on.

For the thousands of projects that currently vendor tokio as a dependency, this is a long-term simplification. The performance is competitive with tokio in early benchmarks - within 5% on throughput - but the real win is the reduced dependency tree. The stabilization timeline is set for Rust 1.90, so the ecosystem has a year to migrate.

PostgreSQL 18 beta adds column-level encryption

PostgreSQL 18 entered beta today, and the headline feature is column-level encryption via a new ENCRYPTED column type. The encryption is transparent to the application - queries still work as normal, but the data is encrypted at rest with keys managed by the server. The implementation uses AEAD with AES-256-GCM, and keys can be rotated without a table rewrite, which is the feature that makes it practical for production.

This doesn’t replace full-disk encryption or TLS, but it addresses the specific case of protecting sensitive columns - PII, tokens, credentials - from database file theft. The beta is available now, and the release candidate is expected in November, with the final release targeted for December. For teams that have been avoiding column-level encryption because of the operational overhead, this version removes the main objection.

The tail: a hardware refresh and a protocol update

NVIDIA announced a mid-cycle refresh of the H200 lineup, adding a 141GB variant with a 12% memory bandwidth increase, priced at $31,000. The upgrade is mostly for inference workloads that are memory-bound rather than compute-bound. Separately, the HTTP/3 working group published an updated draft of the QUIC protocol that adds connection migration improvements for mobile networks. The draft is in last-call review, and the changes are backward-compatible with existing deployments. Both are incremental, but they’re worth tracking if you’re on the hardware or network edge.