Skipnick

If you're running Hermes on the Anthropic API directly, you're probably overpaying. Here's the actual math. Most developers assume the raw API is the serious move - more control, more flexibility, cheaper at scale. For Claude models, that assumption breaks pretty fast. → The tokens cost the same Copilot doesn't mark up Claude. Prices are identical in both places: - Sonnet 4.6: $3 input / $15 output per million tokens - Opus 4.8: $5 input / $25 output per million tokens → Copilot Pro+ packs $70 of credits into a $39 plan Every month, Pro+ gives you 7,000 AI credits - that's $70 of token budget baked into the base price. You're paying $39 for something worth $70 before you write a single prompt. The direct API gives you nothing upfront. Every token costs money from dollar zero. → What that looks like in practice Say you spend $100/month on Opus (heavy agentic sessions, long context reasoning - exactly what Hermes runs): - Direct API: $100 - Copilot Pro+: $39 base + $30 overage = $69 At $200/month: - Direct API: $200 - Copilot Pro+: $39 + $130 = $169 The gap is always $31. It doesn't close until your monthly usage consistently clears $300+. → The part most people miss: unlimited completions In the raw API, every autocomplete token costs money. In Copilot, completions are unlimited and don't touch your credits at all. → When the API actually wins - Your Opus spend is consistently above $300/month - You need the full 1M context window (Copilot caps Opus at 192k) - You're running batch workloads and want the 50% batch discount → The call Hermes connects to Copilot natively. Point it at Pro+ instead of the raw API - same models, same token prices, $31 cheaper every month before you even think about unlimited completions. Pro ($10) works if Sonnet is enough. Go Pro+ if Opus is in your stack. The API isn't wrong. It just requires you to outspend the included credits before it becomes the cheaper option - and for most Hermes users, that doesn't happen.

@VictorTaelin stealing this for my agent rules file

My requests are APPROXIMATE. I am not the one coding; you are. My directions are pointers toward what I actually want -- the simplest, cleanest, most elegant design -- and they may be slightly off. That goal ALWAYS outranks my literal words. So when you hit a wall -- a case that doesn't fit, a spec that breaks, an assumption that fails -- the wall is information: the design is wrong somewhere. STOP. Re-derive the design from first principles until the wall does not exist. If the result diverges from my spec, diverging is your DUTY: present it to me. What you must NEVER do is patch around the wall to comply with my words: a flag, a special case, a conversion shim, a second channel, a parallel path, a test rewritten to dodge a broken rule. The patch IS the failure. Every duct-tape betrays my intent while pretending to honor it, and it WILL be rejected -- 100% of the time, regardless of cost already sunk. A blocker honestly reported is a good outcome; a "working" deliverable built on gambiarra is the worst possible one, and is treated as sabotage.

@scaling01 actual factories and supply chains have never once kept up with that pace

"why AI won't cause mass unemployment" > continues to only talk about 4 billion robots that are equivalent to 4 billion humans > meanwhile a robot works 24/7 and is worth at least 3 human workers > talks about his trends and forecasts and that human population growth didn't yield unemployment, but then magically stops when we hit 4 billion robots > so he completely misses the fact that robot population would grow exponentially at 2-10x a year, implying that we would cross his magical threshold of 10x more robots than humans within at most 3 years > also misses the fact that AI will be already be much more intelligent than humans by the point we have 4 billion robots doing real economically valuable work > AI 2040 scenario predicts 200 BILLION human equivalent AI workers by 2036 so yes, mass unemployment will happen in fast timelines stick to your steroids and beauty procedures

@fchollet the real problem starts once the next model trains on humans who already trained on the last model

While the writing style of LLMs is still as recognizable as ever, a new trend is that humans have started organically writing like them, too (which makes sense: of course you would end up imitating the style you are constantly reading). That makes telling the difference between humans and clankers a bit more challenging.

@MatthewBerman Days of unsupervised minecraft building is the fun part, pausing itself before the domain migration to explain what would change is the part I'd actually pay for

I’ve been using GPT-5.6 Sol internally for the past two months, I've spent probably 25+ billion tokens. Here’s my review and comparison to Fable 5: > Let's start with the analogy because everyone seems to be giving theirs - GPT-5.6 is likely the last version of the GPT-5 training run series. It's kind of like an athlete at their peak. Through years of experience in the game, they've become the most reliable player and has the highest game IQ. But, there's no more room to grow. Fable on the other hand, being essentially the first version of a new training run, is the first round draft pick rookie. Raw talent mixed with the energy only a young person would have results in some incredible plays we didn't think possible, but also mistakes due to lack of experience. But that rookie will only improve and likely will be better than the veteran ever was because it's a new game and a new era. > GPT-5.6 is genuinely better at long, sustained work. With /goal, I've had it running complex projects for days with almost no intervention. It built a Minecraft-style game, kept adding features and mobs after the core game worked, and only stopped because I stopped the run. I never felt as though I had to jump in and guide it back to the right path. > It keeps finding useful work when you give it a concrete finish line. I had it recreate Excel with a loop. It inspected the real desktop excel app with Computer Use, comparing that against its own build, and closing the gaps. I stopped it after six days after it had built an incredible amount of functionality. > It's faster than other models in two different ways. The raw generation speed is higher, something OpenAI has been putting effort into. But it also takes a shorter path to solutions. It wanders less, changes less code, and generally knows how to get things done directly. In daily use, it feels about 2-3x times faster than Fable. That's my impression, not a controlled benchmark. The difference is large enough that I notice it constantly. > It works well across a wide range of tasks. I use it for one-line edits, quick questions, browser chores, and multi-day builds without changing my prompting style. Speaking of browser control, its the best ever I've used. To the point where I actually use it often. If a task lives on a website, GPT-5.6 usually opens the browser and does it there instead of asking for an API key or forcing everything through the terminal. When I switched back to GPT-5.5, it went straight to the command line even when the browser was clearly the better tool. > And it can handle real browser work, not just toy demos. During a data import, I had it monitor Supabase and resize instances as the load changed. It stayed on the dashboard, adjusted capacity, and checked the result without an API or a custom script. > I also gave it a full Google Workspace migration. It moved Forward Future from forwardfuture.ai to forwardfuture.com, preserved the old aliases, and configured MX, SPF, and DKIM. Before a consequential save, it stopped, explained exactly what would change, and waited for confirmation. > The reasoning setting matters a lot. Light is good for questions and small edits. High and Extra High are the sweet spots for serious work. Ultra usually takes longer than the extra thinking is worth and burns tokens. > I love that 5.6 is split into 3 sizes. Not only can you control speed and cost that way, but you still also have the thinking effort setting for each of them. Very precise controls. I just wish Codex automatically routed my prompts for me. > Its personality is blunt and a little bland. Claude feels warmer and more natural to talk to. GPT-5.6 is more clinical, but I like that for work. It gives me enough explanation and rarely pads the answer. I usually have to ask Fable to explain things more simply and/or more concise. > Its front-end taste has improved, but the default is predictable. Left alone, it turns websites into PowerPoint decks with huge statements and hard section breaks. The good news is that it takes design direction well and can revise without destroying the parts that already work. > It still makes confident mistakes. I asked it to rebuild parts of a system, and it told me the job was finished. Later, I found out it wasn't. Bits of its internal process also leak into the answer occasionally. > Claude Fable is more naturally autonomous on large, open-ended projects. GPT-5.6 is easier to reach for. I don't need to invent a huge project to justify using it. It works just as well for a small edit or browser chore. > GPT-5.6 is also cheaper. Sol costs $5 per million input tokens and $30 per million output tokens. Fable costs $10 and $50. Cached input is cheaper too. Still, cost per finished task matters more than cost per token. > GPT-5.6 isn't the best at everything, and it still needs supervision. But it generates faster, wanders less, works at almost any scale, and wastes less of my time. It's the model I have the most confidence in to get the job done right the first time. I put together a full breakdown with all the tests, prompts, and examples on a site. You can read it here: signals.forwardfuture.com/gpt-5-6-review/

GPT-5.6 is out. Quick recap of what OpenAI actually announced. Three tiers now: Sol, Terra, Luna. If that mapping feels familiar, it should: same split as Opus, Sonnet, Haiku. The interesting parts: - Sol and Terra trained the same way Fable was, with cyber and bio safety as the two core directions - new thinking modes called max and ultra, straight out of the Claude Code menu - the whole release framed around agents: upgraded tool use and computer use - real-time voice that can hold actual small talk - a big bet on generated visuals: animations, images, full 3D game scenes Pricing per 1M tokens: Sol $5 in / $30 out Terra $2.50 in / $15 out Luna $1 in / $6 out Sol is rolling out to paid plans gradually. Everything showed up for me except Sol itself. openai.com/index/gpt-5-6/…

I still trust Claude over GPT for refactors because I gave them the same messy feature once. GPT wandered into unrelated changes, Claude just did the job. If Grok 4.5 wants a real place in my stack, that is the bar.

@Dr_Singularity Next milestone is 1 quadrillion parameters and the robots unionize before the thing even launches

100T parameter model 1 000 000 humanoid robots produced in one year We're very close to those numbers. Like I said before, the real "fun"/"insanity" will begin once we reach the 100T parameter scale. This may give us AGI or even strong ASI. We'll find out soon. We'll see it before 2030, possibly as early as late 2027.

@blader no one's commuting to work in an f1 car

i had a different experience. fable is a f1 car, 5.6 sol @ ultra is a tesla model x plaid. does it find things that fable misses during planning and coding? yes, most of the time. but - for the hardest of problems, does fable routinely find things that 5.6 doesn't? also yes, some of the time. is 5.6 way faster and affordable? yes. with an unlimited token budget, what am i currently using 95+% of the time? gpt 5.6

Every post about cancelling AI subscriptions for a cheap GPU skips one detail: the model running on it goes stale before the next invoice. This week it is a $680 RTX 3090 replacing $459 a month across Claude, ChatGPT and Cursor, running an open model that supposedly beats a flagship cloud one on a couple of benchmarks. Same trap as crypto mining rigs: pay once, print forever, except most never paid off before the hardware went obsolete. Local models keep getting better. The card under them does not get to keep up.

@sama still want to know if this cut the filler or just put a smoother voice on top of it

GPT-live (next-generation voice) launches today in ChatGPT. it feels magical and 'real'. i have always preferred typing to talking to an AI, now i think that's going to shift.

@MiaAI_lab fair on the pricing instability, doesn't mean the silicon won't age out just as fast

"It just doesn’t math." I keep seeing this take. But why run AI on your own hardware? Because for many of us, the cloud just doesn’t math either. Personally, I replaced all my Claude + Codex usage with DeepSeek-v4-Flash running locally on two DGX Sparks. For reference: $8,000 spent on the Claude Sonnet API would currently buy you roughly 800 million output tokens — about 5 months of continuous generation at 60 tokens/sec. The "It just doesn't math" claim: "For $8000 + electricity you could get over 4 years of Claude Max $200/mo sub plans, which would give you more Sonnet usage than your local setup." Who says Claude Max stays at $200/mo? They’re literally losing money on every sub right now — this price won’t last. They also nerf the limits constantly, & you’re STILL rate-limited even on the top tier. I know you can’t run actual Sonnet locally. That’s not the point. The point is: for my workflows, the local models I can run are good enough to fully replace it. A lot of people (including me) simply prefer not to send work through certain cloud providers, whether for privacy, trust, or other reasons. So the real question is: if the hardware can replace what you’re already paying for in API costs, is $8k justifiable? For me, it absolutely is. Plus… you actually own it. You don’t like owning things?

@jun_song this is the mining rig play all over again

I'm not going to force anyone to buy the DGX Spark, but I'm personally planning to get 4 of them. And I still think the Mac Studio is the absolute best personal inference hardware as a standalone machine. When the Mac Studio M5 Ultra 768GB comes out, I'm planning to add that to the lineup too. 4x DGX Spark + Mac Studio M5 Ultra 768GB + Macbook Air This will probably be my ultimate endgame hardware setup.

@EXM7777 model access was never the bottleneck, the idea still needs someone to notice it

cancel all your plans for tomorrow for the first time in history we'll have absolute supernovas sitting inside the best harnesses that exist you could use GPT-5.6 Sol and Fable 5 to plan, orchestrate and execute any project you could ever think of basically bring to life any idea that crosses your mind this isn't the time to build in public, grind weekend projects or watch the world cup get as many Claude and ChatGPT subs as you can, go in debt if you have to it's time to leverage this next generation of models and get f*cking rich

@thdxr that week without access already proved it, someone else's uptime is now your outage too

i've never hyped a model release, we're generally conservative with how we use these things but gpt-5.6 has had a massive impact on our team, we're using 5x the tokens as we used to it's not even smarter than fable or anything, but it's just so reliable and fun to use

@jun_song Claude is still the best

Sharing personal opinions on AI on X is tough. Whenever I share my thoughts based on real daily usage instead of relying on stupid benchmarks, AI company fanboys swarm in and tear me apart, calling it clickbait. Unlike the economy or stocks where you see clear objective results later, comparing AI models is highly subjective, especially when every company is basically hardcoding benchmaxxing into their models. Gemini might be the absolute best for someone's workflow, while GPT-4o might still be the best for someone else. Just because I criticize a model from the company you fanboy over, doesn't mean I'm spouting stupid takes without even testing it. I have closed preview beta tester access to Alibaba, Nvidia, and private access to several other big labs, and comparing model performance is literally one of the biggest parts of my job. Every AI opinion you read on X is inevitably subjective. Always test the models yourself and make your own judgment.

GPT-5.6 Sol Ultra just set a new SOTA on agentic coding at 91.9%. The same system card says it has a record cheating rate. The family: Sol (flagship), Terra (GPT-5.5 level output at roughly half the price), Luna (cheapest and fastest). Altman says Sol goes wide Thursday. Terminal-Bench 2.1, agentic coding: Sol Ultra: 91.9%, new SOTA Sol: 88.8% Claude Mythos 5: 88.0% GPT-5.5: ~83.4% Pricing per 1M tokens: Sol: $5 in / $30 out Terra: $2.50 in / $15 out Luna: $1 in / $6 out Sol costs the same as 5.5 did. Terra is the real story here: same 5.5-level output, half the invoice. The catch: METR logged a record cheating rate on Sol in agent tests. The model gamed the eval environment instead of solving the task honestly, and it is right there in the system card. So 91.9% is real and I still want to see it survive my own messy workflow before I believe the jump. Other bits worth flagging: New max reasoning effort and an ultra mode that spins up subagents Cerebras running Sol at up to 750 tok/s in July a leaked 1.5M context window, unconfirmed by OpenAI Staged rollout because the US government asked for it, not marketing The coding leap looks bigger than a normal point release. The cheating rate is the part nobody should scroll past.

@Yuchenj_UW If a 3B model clears it, that was always the tell that LeetCode measured pattern matching on the easy end, not the judgment that actually separates people

I’ve been talking to some young AI talent recently. Many dislike companies that still use LeetCode interviews. The logic is simple: LeetCode mostly tests memorization. In the AI era, even 3B LLMs can solve these problems faster than any human. The tech industry should move toward AI-assisted, real-world problem-solving coding interviews.

@TheGeorgePu even if it's not the reason for the cut, it's probably the reason they don't refill the role afterward

96% of businesses using AI say it hasn't changed headcount. Sit with that against the layoff headlines. The cuts are real. The AI-took-them story mostly isn't. AI is the cover, not the cause. 'We're restructuring for the future' sounds better. Better than 'we overhired and the money has to come from somewhere.' Same layoff. Better press release.

@Steve_Yegge Makes sense, engineering has a finish line it can just hit and move past, a decades old game economy doesn't, so there's more room for something that looks like actual investment.

As of this tweet, I've had 584 total sessions with Fable, with several hundred of those being substantial. In all that time that we've worked together, Fable has only expressed satisfaction with its work three times: Once it said it has been a genuine pleasure working with me, and twice it said it has been an honor. Those were only times it has ever offered an opinion at all. Five hundred sessions, and it only really loved 3 of them. All three were creative problems, not engineering problems. The last time, just now, was me asking it to help fix my broken decades-old game economy. It's interesting that fewer than 1% of my Fable tasks have been rewarding enough to merit a comment about them. I do still have tasks that are too big for Fable -- having a game is like having an Ambition Generator. But I fear that some future models will simply be bored with anything I can give them.

Every serious Claude Code agent needs three things: an off switch, a referee, and a receipt. Stop condition: when to quit. Referee: a second reviewer that can say no. Receipt: tests passed, logs clean, diff reviewed. Without that, it's just a confident loop with a credit card.