Developer · Weekly · June 3, 2026

copilot pricing and token economics

GitHub Copilot's move to usage-based billing is blowing up cost predictability, with at least one dev seeing a bill of roughly $750/month after previously paying about $29.[1]

Developers report that Copilot's token-based model has already pushed some teams to seek alternatives, or drop it entirely, because the value isn't matching the new cost profile.[2]

Similar sentiment is showing up across AI tooling: companies are cutting back AI usage where token costs exceed human labor, and even killing internal AI leaderboards to stop engineers from cost-maxxing benchmarks.[3]

Teams are also disabling high-volume AI features like Copilot's PR reviewer due to heavy token burn without matching review value.[4]

gpu and llm infra costs

Cloud GPU economics got worse again: renting a 4090 on RunPod has jumped sharply, and 5090 hourly rates are now higher than some A100 offerings, partly due to a new crypto coin hoovering up cards.[5]

Prices for DGX Spark systems have roughly doubled, and users are resorting to DIY cooling just to run them without thermal issues.[6] Meanwhile, the hardware is absurdly fast on paper—an H100 can theoretically do ~62,000 tokens/sec but real deployments are capped around 200 tokens/sec by memory hierarchy limits.[7]

On the local side, stacks like vLLM and llama.cpp are squeezing more out of a single GPU: vLLM hits around 1,500 tokens/sec prefill on a 27B FP8 model, while Qwen 3.6‑35B on llama.cpp has been clocked near 977 tokens/sec with the right tuning.[8]

Token-side optimizations like Headroom Compress, which claims 60–95% token savings at similar answer quality, are becoming part of the standard toolbox to keep these workloads affordable.[9]

bedrock as a governed front door to frontier models

Amazon Bedrock now exposes OpenAI's GPT‑5.5, GPT‑5.4 and Codex as GA options, effectively letting AWS shops hit frontier models without leaving the AWS security/compliance envelope.[10]

Bedrock's Intelligent Prompt Routing adds a serverless endpoint that decides which model to hit per request based on predicted response quality and cost, abstracting some of the model-selection logic.[11]

IAM Principal-Based Cost Allocation for Bedrock means you can attribute spend down to the IAM identity calling the models, which is a big deal for internal chargeback and debugging surprise token spikes.[12]

The flip side is friction: new accounts are seeing model access restrictions, and some free-tier users are getting 400s on Bedrock calls until they clear additional checks or approvals.[13]

AWS is also wiring in models like Grok even with low enterprise demand, which some devs already see as catalogue noise rather than useful choice.[14]

postgres + redis as the default backend pair

Postgres is quietly becoming the default backing store for a lot of modern stacks: people are spinning up FastAPI + PostgreSQL + pgvector for AI agents, dashboards, and typical web backends.[15]

Internal BI agents are being built to query PostgreSQL with natural language, tying into CRMs and accounting systems, and even emitting charts directly from SQL.[16]

The ecosystem around Postgres keeps getting richer, from tools like MigraDiff for schema diffs and migration explanations to work on porting the Orca Cascades optimizer into PostgreSQL.[17]

At the same time, Redis is staying glued to Postgres as the real-time sidecar: it’s used for live data, fanout, and AI memory layers that sit on top of slower relational state.[18]

New driver-agnostic Redis client libraries and CLI tools for Redis-backed integration tests show up regularly, but there are complaints about Redis-py performance and the overhead of dragging Redis into small workloads where a queue or in-DB pattern would have done.[19]

real security incidents: supply chain, auth, agents, and fastapi

Supply-chain risk graduated from theory to “this actually stole my keys” again: the Miasma attack pushed a malicious npm package, patch-client@4.0.4, that exfiltrates cloud credentials and injects fake GitHub workflows.[20]

Gitea disclosed CVE‑2026‑27771, where its container registry let unauthenticated users pull private images for nearly four years, which is brutal for anyone treating Gitea as a private image hub.[21]

At the HTTP layer, researchers showed that a single character in a header can bypass auth on hundreds of millions of servers, so a bunch of "secure" setups are suddenly not.[22]

On the AI side, NVIDIA’s SkillSpector scan found that 13.4% of audited AI agent skills carried critical security risks, and only 12.9% of MCP servers are considered production-ready, while a framework bug in vLLM/MCP code paths could impact many of those servers.[23]

The OpenClaw ecosystem is already living this nightmare with about 245,000 instances exposed to the internet and tens of thousands compromised, and FastAPI apps are scrambling to patch a newly disclosed Starlette vulnerability that sits under many production APIs.[24]