Quite a premium for speed. Especially when Gemini 3 Pro is 1.8x the tokens/sec speed (of regular-speed Opus 4.6) at 0.45x the price [2]. Though it's worse at coding, and Gemini CLI doesn't have the agentic strength of Claude Code, yet.
no, it’s for Max subscribers to enable “use API when running out of session limit”. the assumption (probably) being that many will forget to turn it off, and they’ll earn it back that way.
This was my first thought, but by default, you have no automatic reload of your prepaid account. Which I think is for once user friendly. They could have applied a dark pattern here.
A useful feature would be slow-mode which gets low cost compute on spot pricing.
I’ll often kick off a process at the end of my day, or over lunch. I don’t need it to run immediately. I’d be fine if it just ran on their next otherwise-idle gpu at much lower cost that the standard offering.
Still do. Great for workloads where it's okay to bundle a bunch of requests and wait some hours (up to 24h, usually done faster) for all of them to complete.
Yep same, I often think why this isn’t a thing yet. Running some tasks in the night at e.g. 50% of the costs - there’s the batch api but that is not integrated in e.g. claude code
> I’ll often kick off a process at the end of my day, or over lunch. I don’t need it to run immediately. I’d be fine if it just ran on their next otherwise-idle gpu at much lower cost that the standard offering.
If it's not time sensitive, why not just run it at on CPU/RAM rather than GPU.
But it's incredibly incapable compared to SOTA models. OP wants high quality output but doesn't need it fast. Your suggestion would mean slow AND low quality output.
I'm assuming GP means 'run inference locally on GPU or RAM'. You can run really big LLMs on local infra, they just do a fraction of a token per second, so it might take all night to get a paragraph or two of text. Mix in things like thinking and tool calls, and it will take a long, long time to get anything useful out of it.
I’ve been experimenting with this today. I still don’t think AI is a very good use of my programming time… but it’s a pretty good use of my non-programming time.
I ran OpenCode with some 30B local models today and it got some useful stuff done while I was doing my budget, folding laundry, etc.
It’s less likely to “one shot” apples to apples compared to the big cloud models; Gemini 3 Pro can one shot reasonably complex coding problems through the chat interface. But through the agent interface where it can run tests, linters, etc. it does a pretty good job for the size of task I find reasonable to outsource to AI.
This is with a high end but not specifically AI-focused desktop that I mostly built with VMs, code compilation tasks, and gaming in mind some three years ago.
Yes, this is what I meant. People are running huge models at home now, I assumed people could do it on premises or in a data center if you're a business, presumably faster... but yeah it definitely depends on what time scales we're talking.
I'd love to know what kind of hardware would it take to do inference at the speed provided by the frontier model providers (assuming their models were available for local use).
Huge models? First you have to spend $5k-$10k or more on hardware. Maybe $3k for something extremely slow (<1 tok/sec) that is disk-bound. So that's not a great deal over batch API pricing for a long, long time.
Also you still wouldn't be able to run "huge" models at a decent quantization and token speed. Kimi K2.5 (1T params) with a very aggressive quantization level might run on one Mac Studio with 512GB RAM at a few tokens per second.
To run Kimi K2.5 at an acceptable quantization and speed, you'd need to spend $15k+ on 2 Mac Studios with 512GB RAM and cluster them. Then you'll maybe get 10-15 tok/sec.
How much extra power do you think you would need to run an LLM on a CPU (that will fit in RAM and be useful still)? I have a beefy CPU and if I ran it 24/7 for a month it would only cost about $30 in electricity.
Note that you can't use this mode to get the most out of a subscription - they say it's always charged as extra usage:
> Fast mode usage is billed directly to extra usage, even if you have remaining usage on your plan. This means fast mode tokens do not count against your plan’s included usage and are charged at the fast mode rate from the first token.
Although if you visit the Usage screen right now, there's a deal you can claim for $50 free extra usage this month.
So it's basically useless then. Even with Claude Max I have to manage my usage when doing TDD, and using ccusage tool I've seen that I'd frequently hit $200 per day if I was on the API. At 6x cost you'll burn through $50 in about 20 minutes. I wish that was hyperbole.
Because we all prefer it over Gemini and Codex. Anthropic knows that and needs to get as much out of it as possible while they can. Not saying the others will catch up soon. But at some point other models will be as capable as Opus and Sonnet are now, and then it's easier to let price guide the choice of provider.
Counterintuitively, I feel like this will not be super useful, at least for me. My bottleneck is MY ability to parse and understand LLM-generated code. The agent can code a lot faster than I can read and understand its output.
Smart if you can get away with it. For non trivial things you can't quite get away with it.
I've started asking Claude to make me a high-level implementation plan and basically prompt ME to write it. For the most part I walk through it and ask Claude to just do it. But then 10% of the time there is a pretty major issue that I investigate, weigh pros/cons, and then decide to change course on.
Those things accumulate. Maybe 5-10 things over the course of an MVP that I wouldn't really have a clue about if I let Claude just dutiful implement it's own plan.
Spend time building a test harness and evaluations of whether the solution meets the requirements. Then you don't need to look at the code because those other pieces will bring the necessary guarantees and trust.
I was thinking about inhouse model inference speeds at frontier labs like Anthropic and OpenAI after reading the "Claude built a C compiler" article.
Having higher inference speed would be an advantage, especially if you're trying to eat all the software and services.
Anthropic offering 2.5x makes me assume they have 5x or 10x themselves.
In the predicted nightmare future where everything happens via agents negotiating with agents, the side with the most compute, and the fastest compute, is going to steamroll everyone.
LLM APIs are tuned to handle a lot of parallel requests. In short, the overall token throughput is higher, but the individual requests are processed more slowly.
The scaling curves aren't that extreme, though. I doubt they could tune the knobs to get individual requests coming through at 10X the normal rate.
This likely comes from having some servers tuned for higher individual request throughput, at the expense of overall token throughput. It's possible that it's on some newer generation serving hardware, too.
This makes no sense. It's not like they have a "slow it down" knob, they're probably parallelizing your request so you get a 2.5x speedup at 10x the price.
All of these systems use massive pools of GPUs, and allocate many requests to each node. The “slow it down” knob is to steer a request to nodes with more concurrent requests; “speed it up” is to route to less-loaded nodes.
But it’s actually not so difficult is it? The simplest way to make a slow pool is by having fewer GPUs and queuing requests for the non-premium users. Dead simple engineering.
Oh, of course. That’s just conspiratorial thinking. Paying to be in a premium pool makes sense, all of this “they probably serve rotten food to make people pay for quality food” nonsense is just silly.
What they are probably doing is speculative decoding, given they've mentioned identical distribution at 2.5x speed. That's roughly in the range you'd expect for that to achieve; 10x is not.
It's also absolute highway robbery (or at least overly-aggressive price discrimination) to charge 6x for speculative decoding, by the way. It is not that expensive and (under certain conditions, usually very cheap drafter and high acceptance rate) actually decrease total cost. In any case, it's unlikely to be even a 2x cost increase, let alone 6x.
Where on earth are you getting these numbers?
Why would a SaaS company that is fighting for market dominance withhold 10x performance if they had it? Where are you getting 2.5x?
This is such bizarre magical thinking, borderline conspiratorial.
There is no reason to believe any of the big AI players are serving anything less than the best trade off of stability and speed that they can possibly muster, especially when their cost ratios are so bad.
Thats also called slowing down default experience so users have to pay more for the fast mode. I think its the first time we are seeing blatant speed ransoms in the LLMs.
That's not how this works. LLM serving at scale processes multiple requests in parallel for efficiency. Reduce the parallelism and you can process individual requests faster, but the overall number of tokens processed is lower.
They have contracts with companies, and those companies wont be able to change quickly. By the time those contracts will come back for renewals it will already be too late, their code becoming completely unreadable by humans. Individual devs can move quickly but companies don't.
Are you at all familiar with the architecture of systems like theirs?
The reason people don't jump to your conclusion here (and why you get downvoted) is that for anyone familiar with how this is orchestrated on the backend it's obvious that they don't need to do artificial slowdowns.
Seriously, thinking at the price structure of this (6x the price for 2.5x the speed, if that's correct) it seems to target something like real time applications with very small context. Maybe vocal assistants? I guess that if you're doing development it makes more sense to parallelize over more agents rather than paying that much for a modest increase in speed.
I’m curious what’s behind the speed improvements. It seems unlikely it’s just prioritization, so what else is changing? Is it new hardware (à la Groq or Cerebras)? That seems plausible, especially since it isn’t available on some cloud providers.
Also wondering whether we’ll soon see separate “speed” vs “cleverness” pricing on other LLM providers too.
It comes from batching and multiple streams on a GPU. More people sharing 1 GPU makes everyone run slower but increases overall token throughput.
Mathematically it comes from the fact that this transformer block is this parallel algorithm. If you batch harder, increase parallelism, you can get higher tokens/s. But you get less throughput. Simultaneously there is also this dial that you can speculatively decode harder with fewer users.
Its true for basically all hardware and most models. You can draw this Pareto curve of how much throughput per GPU vs how many tokens per second per stream. More tokens/s less total throughput.
See this graph for actual numbers:
Token Throughput per GPU vs. Interactivity
gpt-oss 120B • FP4 • 1K / 8K • Source: SemiAnalysis InferenceMAX™
> If you batch harder, increase parallelism, you can get higher tokens/s. But you get less throughput. Simultaneously there is also this dial that you can speculatively decode harder with fewer users.
I think you skipped the word “total throughout” there right? Cause tok/s is a measure of throughput, so it’s clearer to say you increase throughput/user at the expense of throughput/gpu.
I’m not sure about the comment about speculative decode though. I haven’t served a frontier model but generally speculative decode I believe doesn’t help beyond a few tokens, so I’m not sure you can “speculatively decode harder” with fewer users.
There are a lot of knobs they could tweak. Newer hardware and traffic prioritisation would both make a lot of sense. But they could also lower batching windows to decrease queueing time at the cost of lower throughput, or keep the KV cache in GPU memory at the expense of reducing the number of users they can serve from each GPU node.
2.4x faster memory - which is exactly what they are saying the speedup is. I suspect they are just routing to GB200 (or TPU etc equivalents).
FWIW I did notice _sometimes_ recently Opus was very fast. I put it down to a bug in Claude Code's token counting, but perhaps it was actually just occasionally getting routed to GB200s.
Dylan Patel did analysis that suggests lower batch size and more speculative decoding leads to 2.5x more per-user throughput for 6x the cost for open models [0]. Seems plausible this could be what they are doing. We probably won't get to know for sure any time soon.
Regardless, they don't need to be using new hardware to get speedups like this. It's possible you just hit A/B testing and not newer hardware. I'd be surprised if they were using their latest hardware for inference tbh.
Why does this seem unlikely? I have no doubt they are optimizing all the time, including inference speed, but why could this particular lever not entirely be driven by skipping the queue? It's an easy way to generate more money.
Yes it's 100% prioritization. Through that it's also likely running on more GPUs at once but that's an artifact of prioritization at the datacenter level. Any task coming into an AI datacenter atm is split into fairly fined grained chunks of work and added to queues to be processed.
When you add a job with high priority all those chunks will be processed off the queue first by each and every GPU that frees up. It probably leads to more parallelism but... it's the prioritization that led to this happening. It's better to think of this as prioritization of your job leading to the perf improvement.
Here's a good blog for anyone interested which talks about prioritization and job scheduling. It's not quite at the datacenter level but the concepts are the same. Basically everything is thought of as a pipeline. All training jobs are low pri (they take months to complete in any case), customer requests are mid pri and then there's options for high pri. Everything in an AI datacenter is thought of in terms of 'flow'. Are there any bottlenecks? Are the pipelines always full and the expensive hardware always 100% utilized? Are the queues backlogs big enough to ensure full utilization at every stage?
Amazon Bedrock has a similar feature called "priority tier": you get faster responses at 1.75x the price. And they explicitly say in the docs "priority requests receive preferential treatment in the processing queue, moving ahead of standard requests for faster responses".
I wonder if they might have mostly implemented this for themselves to use internally, and it is just prioritization but they don't expect too many others to pay the high cost.
I’d love to hear from engineers who find that faster speed is a big unlock for them.
The deadline piece is really interesting. I suppose there’s a lot of people now who are basically limited by how fast their agents can run and on very aggressive timelines with funders breathing down their necks?
> I’d love to hear from engineers who find that faster speed is a big unlock for them.
How would it not be a big unlock? If the answers were instant I could stay focused and iterate even faster instead of having a back-and-forth.
Right now even medium requests can take 1-2 minutes and significant work can take even longer. I can usually make some progress on a code review, read more docs, or do a tiny chunk of productive work but the constant context switching back and forth every 60s is draining.
I won't be paying extra to use this, but Claude Code's feature-dev plugin is so slow that even when running two concurrent Claudes on two different tasks, I'm twiddling my thumbs some of the time. I'm not fast and I don't have tight deadlines, but nonetheless feature-dev is really slow. It would be better if it were fast enough that I wouldn't have time to switch off to a second task and could stick with the one until completion. The mental cost of juggling two tasks is high; humans aren't designed for multitasking.
Hmm I’ve tried two modes: one is to stay focused on the task at hand, but spin up alternative sessions to do documentation, check alternative hypotheses, second-guess things the main session is up to. — The other is to do an unrelated task in another session. I find this gets more work done in a day but is exhausting. With better scaffolding and longer per-task run times (longer tasks in the METR sense), could be more sustainable as a manager of agents.
3-4 parallel projects is the norm now, though I find task-parallelism still makes overlap reduction bothersome, even with worktrees. How did you work around that?
it's simpler than that - making it faster means it becomes less of an asynchronous task.
current speeds are "ask it to do a thing and then you the human need find something else to do for minutes (or more!) while it works". at a certain point at it being faster you just sit there and tell it to do a thing and it does and you just constantly work on the one thing.
cerebras is just about fast enough for that already, with the downside of being more expensive and worse at coding than claude code.
it feels like absolute magic to use though.
so, depends how you price your own context switches, really.
The only time I find faster speed to be a big unlock is when iterating on UI stuff. If you're talking to your agent, with hot reload and such the model can often be the bottleneck in a style tuning workflow by a lot.
So us normal pro accounts have slow mode. Thanks anthropic.
I'm currently testing Kimi2.5 with cli, works great and fast. Even comes with a web interface so you can communicate with you kimi-cli instance (remote even if you use vpn).
Just when you thought it was safe to use Opus 4.5 at 1/3 the cost, they go and add a 6x 'bank-breaking mode' - So now accidental bankruptcy is just one toggle away.
My (and many others) normal workflow includes a planning phase, followed by an implementation phase. For me the most useful time for fast mode would be during that planning phase.
The current "clear context and execute plan" would be great to be joined by a, "clear context, switch to regular speed mode, and execute plan".
I even think I would not require fast mode for the explore agents etc - they have so much to do that I accept that takes a while. being able to rapidly iterate on the plan before setting it going would make it easier.
A developer can blast millions of tokens in minutes. When you have a context size of 250k that’s just 4 queries. But with tool usage and subsequent calls etc it can easily just do many millions in one request
But if you just ask a question or something it’ll take a while to spend a million tokens…
Yeah that’s what they try to do with the latest coding agents sub agents which only have the context they need etc. but atm it’s too much work to manage contexts at that level
I use one Claude instance at a time, roughly fulltime (writes 90% of my code). Generally making small changes, nothing weird. According to ccusage, I spend about $20 of tokens a day, a bit less than 1 MTOK output tokens a way. So the exact same workflow would be about $120 for higher speed.
The API price is 6x that of normal Opus, so look forward to a new $1200/mo subscription that gives you the same amount of usage if you need the extra speed.
The writing has been on the wall since day 1. They wouldn't be marketing a subscription being sold at a loss as hard as they are if the intention wasn't to lock you in and then increase the price later.
What I expect to happen is that they'll slowly decrease the usage limits on the existing subscriptions over time, and introduce new, more expensive subscription tiers with more usage. There's a reason why AI subscriptions generally don't tell you exactly what the limits are, they're intended to be "flexible" to allow for this.
It's explicitly called out as excluded in the blue info bubble they have there.
> Fast mode usage is billed directly to extra usage, even if you have remaining usage on your plan. This means fast mode tokens do not count against your plan’s included usage and are charged at the fast mode rate from the first token.
This is gold for Anthropic's profitability. The Claude Code addicts can double their spend to plow through tokens because they need to finish something by a deadline. OpenAI will have a similar product within a week but will only charge 3x the normal rate.
This angle might also be NVidias reason for buying Groq. People will pay a premium for faster tokens.
I switched back to 4.5 Sonnet or Opus yesterday since 4.6 was so slow and often “over thinking” or “over analyzing” the problem space. Tasks which accurately took under an minute in Sonnet 4.5 were still running after 5 minutes in 4.6 (yeah I had them race for a few tasks)
Someone of this could be system overload I suppose.
Edit ~/.claude/settings.json and add "effortLevel": "medium". Alternatively, you can put it in .claude/settings.json in a project if you want to try it out first.
They recommend this in the announcement[1], but the way they suggest doing it is via a bogus /effort command that doesn't exist. See [2] for full details about thinking effort. It also recommends a bogus way to change effort by using the arrow keys when selecting a model, so don't use that either.
Pathetic how they have no support for modifying sampling settings, or even a "logit_bias" so I can ban my claude from using the EM dash (and regular dash), semicolons, or "not". Also will upweight things like exclamation points
Clearly those whose job it is to "monitor" folks use this as their "tell" if someone AI generated something. That's why every major LLM has this particular slop profile. It's infuriating.
Good to know it works for some people! I think it's another issue where they focus too much on MacOS and neglect Windows and Linux releases. I use WSL for Claude Code since the Windows release is far worse and currently unusable do to several neglected issues.
Hoping to see several missing features land in the Linux release soon.
I'm also feeling weak and the pull of getting a Mac is stronger. But I also really don't like the neglect around being cross-platform. It's "cross-platform" except a bunch of crap doesn't work outside MacOS. This applies to Claude Code, Claude Desktop (MacOS and Windows only - no Linux or WSL support), Claude Cowork (MacOS only). OpenAI does the same crap - the new Codex desktop app is MacOS only. And now I'm ranting.
I'm on v2.1.37 and I have it set to auto-update, which it does. I also tend to run `claude update` when I see a new release thread on Twitter, and usually it has already updated itself.
Claude Code CLI 2.1.39 released a few hours ago fixes the problem. They didn't note it in the changelog though. Seems like a significant bug fix. ¯\_(ツ)_/¯
Yep, and their documentation AI assistant will egregiously hallucinate whatever it thinks you want to hear, then repeat itself in a loop when you tell it that it's wrong.
Yesterday I asked a question about a Claude Code setting inside Claude Code, don't recall which, and their builtin documentation skill—something like that—ended up doing a web search and found a wrong answer on a third party site. Later I went to their documentation site and it was right there in the docs. Wonder why they can't bundle an AI-friendly version of their own docs (can't be more than a few hundred KBs compressed?) inside their 174MB executable.
It's insane that they concluded the builtin introspection skill for claude documentation should do a web search instead of simply packing the correct documentation in local files. I had the same experience like you, wasting tokens and my time because their architecture decision doesn't work in practice.
I have to google the correct Anthropic documentation and pass that link to claude code because claude isn't able to do the same reliably in order to know how to use its own features.
They used to? I have a distinct memory of it doing exactly that a few months ago. Maybe it got dropped in the mad dash that passes for CC sprint cycles
Yeah, nothing is sped up, their initial deployment of 4.6 is so unbearably slow they are just now offering you the opportunity to pay more for the same experience of 4.5. What's the word for that?
Honestly, Open AI isn't worth it. I cancelled my Open AI plan (and hopefully will delete my account soon once I export all my data out) because of philosophy differences. They shared they are evaluating a model where they can get a % of your business in exchange for letting you use code generated by their AI models. That and the possible advertising angle. But, that's not even the worst, I asked ChatGPT to fairly evaluate the risky model where one for profit corporation holds your entire intimate personal details and uses it for advertising, it staunchly defended OpenAI. That was the nail in the coffin for me.
Contrast to this - Anthropic actually asks you if you want their AI to remember details about you and they have lot of toggles around privacy. I don't care if they make money from extra tokens as long as they don't go the Open AI route.
> They shared they are evaluating a model where they can get a % of your business in exchange for letting you use code generated by their AI models.
That's a gross mischaracterization of what the CFO said. She basically just said the pricing space is huge, and they've even explored things like royalty models.
I'm guessing you just saw a headline and read nothing into it.
Isn't it what "evaluating a model where they can get a % of your business in exchange for letting you use code generated by their AI models" precisely mean?
If they find that this business model is most profitable for OpenAI, and that they can somehow release models better than any competitor, wouldn't they say they want royalties ? That's what Unity (the game engine) does so it wouldn't be unseen.
"As intelligence moves into scientific research, drug discovery, energy systems, and financial modeling, new economic models will emerge. Licensing, IP-based agreements, and outcome-based pricing will share in the value created. That is how the internet evolved. Intelligence will follow the same path."
"Intelligence will follow the same path."
This is from their official press release. Also, when you talk about "royalty models", what exactly do you think it means?
> Now they have a profit motive for slowing down the normal service.
Sure. But for now, this is a competitive space. The competitors offer models at a decent quality*speed/price ratio and prevent Anthopic from going too far downhill.
Actually, as I think about it... I don't enjoy any other model as much as Opus 4.5 and 4.6. For me, this is no longer a competitive space. Anthropic are in full right to charge premium prices for their premium product.
The difference being that Airlines and food delivery did make a profit, just figured they had to do these tricks to earn some more. Mature businesses resort to lowering quality, fake scarcity.
Here the scarcity is real, and profits are nowhere to be seen
These schemes will soon fall apart entirely when an open weight model can run on Groq/Cerebras/SambaNova at even higher speeds and be just fine for all tasks. Arguably already the case, but not many know yet.
Yes, but GPT-5.2 and Codex were widely considered slower than Opus before that. They still feel very slow, at least on high. I should give medium a try more often.
Given how little most of us can know about the true cost of inference for these providers (and thus the financial sustainability of their services), this is an interesting signal. Not sure how to interpret it, but it doesn’t feel like it bodes well.
Given that providers of open source models can offer Kimi K2.5 at input $0.60 and output $2.50 per million tokens, I think the cost of inference must be around that. We would still need to compare the tokens per second.
The most interesting part about this is that Anthropic's own employees are probably using this 24/7 to develop the next model and it's barely affordable to anyone else.
The async AI + sync approval bottleneck is real. One thing that helped me: stop being physically desk-bound during those wait times.
I built ForkOff to solve this - when Claude needs approval, push notification to your phone, one-tap approve from anywhere. Turns out you don't actually need to be at your desk for most approvals.
The fast mode helps with speed, but even faster is letting the AI work while you're literally anywhere else. Early access: forkoff.app
(And yes, the pricing for fast mode is wild - $100 burned in 2 hours per some comments here!)
> ...stop being physically desk-bound during those wait times.
> I built ForkOff to solve this
This does sound useful but I have to laugh because I just work out or play with my children. If you enjoy calisthenics and stretching its great to use Claude while not feeling chair-bound. Programming becomes more physical!
AI trained on real time data will always and only get dumber over time. Reversion to mean of human IQ. just like every web forum ever. Eternal September.
That’s why gen 1-3 AI felt so smart. It was trained on the best curated human knowledge available. And now that’s done it’s just humanities brain dumps left to learn from.
2 ways out: self referential learning gen 1-3 AIs. Or, Pay experts to improve datasets and not training with general human data. Inputs and outputs.
Could be a use for the $50 extra usage credit. It requires extra usage to be enabled.
> Fast mode usage is billed directly to extra usage, even if you have remaining usage on your plan. This means fast mode tokens do not count against your plan’s included usage and are charged at the fast mode rate from the first token.
After exceeding the increasingly shrinking session limit with Opus 4.6, I continued with the extra usage only for a few minutes and it consumed about $10 of the credit.
I can't imagine how quickly this Fast Mode goes through credit.
While it's an excellent way to make more money in the moment, I think this might become a standard no-extra-cost feature in several months (see Opus becoming way cheaper and a default model within months). Mental load management while using agents will become even more important it seems.
Why would they cut a money making feature? In fact I am already imagining them asking for speed ransom every time you are in a pinch, some extra context space will also become buyable. Anthropic is in a penny pincher phase right now and they will try to milk everything. Watch them add micro transactions too.
If models really do continue to get more expensive then it's not going to make sense to let everyone at your org have equal budget for spend. We're on track for a world where there are the equivalent of F1 drivers for ai.
One persons output doesn't scale for an entire org no matter how expensive and good the AI is. It's either good enough that anyone can do it, or bad enough that a human needs to be in control which caps the output at human understanding. It will always be more efficient to have every engineer be boosted a little bit than a single one a lot.
I don't think this is the case, according to the docs, right? The effort level will use fewer tokens, but the independent fast mode just somehow seems to use some higher priority infrastructure to serve your requests.
You're comparing two different things. It's not useless knowledge, it's something you need to understand.
Opus fast mode is routed to different servers with different tuning that prioritizes individual response throughput. Same model served differently. Same response, just delivered faster.
The Gemini fast mode is a different model (most likely) with different levels of thinking applied. Very different response.
Mintlify is the best example of a product that is just nice. They don't claim to have a moat, or weird agi vibes, or whatever. It just works and it's pretty. 10m arr right there
AI data centers are a whole lot of pipelines pumping data around utilizing queues. They want those expensive power hungry cards near 100% utilized at all times. So they have a queue of jobs on each system ready to run, feeding into the GPU memory as fast as completed jobs are read out of memory (and passed into the next stage) and they aim to have enough backlog in these queues to keep the pipeline full. You see responses in seconds but at the data center you're request was broken into jobs, passed around into queues, processed in an orderly manner and pieced back together.
With fast mode you're literally skipping the queue. An outcome of all of this is that for the rest of us the responses will become slower the more people use this 'fast' option.
I do suspect they'll also soon have a slow option for those that have Claude doing things overnight with no real care for latency of the responses. The ultimate goal is pipelines of data hitting 100% hardware utilization at all times.
Hmm not sure I agree with you there entirely. You're right there's queues to ensure that you max out the hardware with concurrent batches to _start_ inference, but I doubt you'd want to split up the same job into multiple bits and move them around servers if you could at all avoid it.
It requires a lot of bandwidth to do that and even at 400gbit/sec it would take a good second to move even a smaller KV cache between racks even in the same DC.
Will this mean that when cost is more important than latency that replies will now take longer?
I’m not in favor of the ad model chatgpt proposes. But business models like these suffer from similar traps.
If it works for them, then the logical next step is to convert more to use fast mode. Which naturally means to slow things down for those that didn’t pick/pay for fast mode.
We’ve seen it with iPhones being slowed down to make the newer model seem faster.
Not saying it’ll happen. I love Claude. But these business models almost always invite dark patterns in order to move the bottom line.
Inference is run on shared hardware already, so they're not giving you the full bandwidth of the system by default. This most likely just allocates more resources to your request.
LLM programming is very easy. First you have to prompt it to not mistakes. Then you have to tell it to go fast. Software engineering is over bro, all humans will be replaced in 6 days bro
This pricing is pathetic. I've been using it for two hours at what I consider "normal" interactive speed and it burned $100. Normally the $200 subscription is enough for an entire month. I guess if you are rich, you can pay 40 times as much for roughly double speed (assuming 8 hours usage a day, 5 days a week)?
Edit: I just realized that's with the currently 50% discounted price! So you can pay 80 times as much!
Smart business model decision, since most people and organizations prefer regular progress.
In the future this might be the reason enterprise software companies win - because they can use their customer funds to pay for faster tokens and adaptions.
I redeemed my 50 USD credit to give it a go. In literally less than 10 minutes I spent 10 USD. Insane. I love Claude Code, but this pricing is madness.
The title is doing a lot of heavy lifting. No credible scientist describes AMOC collapse as the “end of humanity.” It would be catastrophic, full stop: European winters dropping 5-15°C, massive disruptions to global food production, coastal flooding, monsoon shifts affecting billions. But humans survived the last AMOC shutdown during the Younger Dryas roughly 12,900 years ago with stone tools and no supply chains. We’d suffer enormously. We wouldn’t go extinct. Framing it that way makes it easier for skeptics to dismiss the real risk.
Quite a premium for speed. Especially when Gemini 3 Pro is 1.8x the tokens/sec speed (of regular-speed Opus 4.6) at 0.45x the price [2]. Though it's worse at coding, and Gemini CLI doesn't have the agentic strength of Claude Code, yet.
[1] - https://x.com/claudeai/status/2020207322124132504 [2] - https://artificialanalysis.ai/leaderboards/models