OAI really trying to push folks to either the ad infested tier for gen chat (ChatGPT Go) or Pro ($100/month) for code.
I guess the $20 Plus plan is the red-headed step child now.
Meanwhile, Z.ai coder is what…$9/month? Hmmm.
I don’t have a local rig powerful enough to natively run a good coder (Qwen 3 next), but “I am altering the deal, pray I do not alter it any further” is THE perfect use case for running your shit locally.
PS: It looks like Pro and Plus tiers have gone up a smidge too? Eg: last I checked they were $20 and $100; they now show $30 and $155. I cancelled my account so OAI gifted me a free month (lol) but seriously…fuck that.
It’s not entirely unexpected, all the AI companies have been heavily subsidising inference to get customers.
I don’t use Codex but I’ve been experimenting with ECA and I can track my token API costs across Gemini and Anthropic. I’m mostly using Gemini and a heavy days usage would be £1.50 in API costs and I’m certainly not doing that every day. I have to wonder if these Codex users are conscious of how many tokens they are burning underneath or just YOLOing everything until the computer says no?
ECA allows you to mix and match models to sub-agents and I could certainly see me offloading some tasks like code exploration to a locally hosted models and saving the expensive reasoning tokens for planning.
PS: Just to let you know - ECA is fucking amazing! Thank you VERY much for that. I’m currently using Sonnet 4.6 via OpenRouter API to boss around my local Qwen3-4… so far so good :)
Right now, I can see -
Sonnet 4.6 API is not too bad at all at moment
- 1,000,000 context $3/M input tokens / $15/M output tokens
GLM 5.1 is STUPID cheap for what it is (promo?)
- 202.8K context $1.395/M input tokens / $4.40/M output tokens
And Gemma 4 31B is just taking the piss
- 262.1K context $0.14/M input tokens / $0.40/M output tokens
(or FREE via Google…but you know…ZDR)
Absolutely true. If I had to pull a number out of thin air, I’d say they were still probably under-charging what it actually costs them to run these things by an order of magnitude or two. So right now, Codex pro costs $150…but in a year or two? $300-400 or even $500? I can see them slowly ratcheting it up. It’s the same old story we’ve seen played out before (eg: Uber, Spotify, Netflix etc).
Doesn’t mean it’s one we should particularly want to see repeat tho.
Like you, I like the notion of mixing and matching local agents for grunt work and off-loading the thinking to API or SOTA. I hadn’t heard of ECA - that looks like it’s right up my alley. Thanks for that
How is Gemini 4 open source for coding? Seems to run well even on mid tier gpu’s
I genuinely haven’t tried it. Can someone here comment on it?
PS: I take it you mean Gemma not Gemini?
wow I love that my brain has been autocorrecting Gemma to Gemini for the last week
You and everyone else :)
I asked Claude btw. Here’s what the Frenchman thinks (yes, I know, Claude Shannon was not French)
Great question — Gemma 4 is a genuinely impressive release. Let me break it down by model with a coding-specific score relative to your Opus 4.6 = 10/10 benchmark.
The lineup — Gemma 4 comes in four sizes: E2B (2.3B effective parameters), E4B (4.5B effective), a 31B dense model, and a 26B MoE with only 4B active parameters. All are Apache 2.0 licensed and run locally.
Coding benchmarks (the numbers that matter here):
On LiveCodeBench v6, the 31B scores 80.0%, the 26B MoE 77.1%, the E4B 52.0%, and the E2B 44.0% — compared to just 29.1% for the previous Gemma 3 27B. The Codeforces ELO tells an even sharper story: the 31B reaches 2150 ELO, the 26B MoE 1718, the E4B 940, and the E2B just 633.
My scores vs. Opus 4.6 = 10/10 for local coding:
Gemma 4 31B — ~7.5/10 This is the headline act. An 80% LiveCodeBench and Codeforces ELO of 2150 is elite for an open-weights model. The catch: at 31B dense, you’ll need ~20GB VRAM quantized (Q4), so it needs a beefy local setup (e.g. a 24GB GPU or Apple Silicon M2 Ultra+). If your hardware can handle it, this genuinely punches into serious territory.
Gemma 4 26B A4B (MoE) — ~7/10 The most exciting model for local use IMO. It reaches an estimated LMArena score of 1441 with just 4B active parameters, meaning it runs fast with the memory footprint of a ~4B model while delivering near-31B quality. On a 16GB Mac or a mid-range GPU, this is a remarkable coding companion. The 77.1% LiveCodeBench is only fractionally behind the 31B.
Gemma 4 E4B — ~5/10 A solid coder for something running in ~6GB RAM. 52% LiveCodeBench won’t blow your mind on hard problems, but for autocomplete, boilerplate, explanations, and routine scripts it’s genuinely useful. The E4B also supports audio input alongside image and text, making it unusually versatile for its size.
Gemma 4 E2B — ~3.5/10 Impressive that it works at all at ~2.3B effective parameters, but the 44% LiveCodeBench and ELO of 633 show real limits on complex coding tasks. Fine for quick snippets or code explanation on very constrained hardware (Raspberry Pi, etc.), but not a serious daily driver for development.
The big picture: The 26B MoE is probably the most exciting model in the lineup for most people — it’s the one that truly shifts what’s possible locally. It runs with first-class support in llama.cpp, MLX, and mistral.rs, so dropping it into tools like LM Studio, Jan, or a local agent is straightforward. For Apple Silicon users especially, the MoE model via MLX with TurboQuant quantization should be very fast. Compared to Opus 4.6 it still falls short on deep reasoning, multi-file refactoring, and architectural planning — but for a model running entirely on your machine with no API costs, the gap has genuinely narrowed.
Blah blah blag…sounds good…but def something you’d need to validate yourself. I’m more tempted by the Qwen models and have been looking at the Yi ones as well…though last time I tested the Yi-9B coder in 2025 it went…badly LOL.
That MoE might be worth a shot though.
In case you don’t want to deal with Shitter -
