Promotional pricing currently mitigates the model's high token costs
Sonnet 5.0 just dropped. Early vibes
> huge token guzzler, promotion pricing helps > step above 4.6, but not really as good as Opus 4.8
Opus 4.8 still seems better than Sonnet 5 in terms of cost/performance trade-off.
Will confirm on LiveBench shortly
Many users criticize Sonnet 5.0 as a token-guzzling downgrade from Opus 4.8 that only seems viable under temporary promo pricing, while some praise its self-correcting agent performance.
No Digg Deeper questions have been answered for this story yet.
@bindureddy Sonnet 5 solid, but Opus 4.8 still hits different
@bindureddy Token guzzler? Is that from a benchmark?
@bindureddy If you want to get anything half decent, xhigh or max. Turns out it costs more than Opus 4.8 if you do that! Pointless model
@bindureddy i mean if promotion pricing helps then 5.0 is just a rental deal compared to opus 4.8 muscle
@bindureddy Waiting for the coders to test it but my hunch is that it’s essentially “better” than Opus 4.7/4.8
Fable is the sucessor of 4.6 and another level of intelligence.
@bindureddy awaiting benchmark tests
@bindureddy the token burn concern is real
new tokenizer eating 35% more per prompt is the footnote most people are skipping
@bindureddy Sonnet 5.0's token guzzling suggests inefficiency despite pricing; how does sparsity actually reduce GPU memory?
@bindureddy Opus 4.8 still holding the crown? 👀
@bindureddy token guzzler with promo pricing is just every new model's origin story, ask again in 3 months when the discount ends lol
@bindureddy Story of my life these days. Opus 4.8 still wins on ratio. By the time enterprise procurement approves Sonnet 5.0, there'll be three more models to re-evaluate. The benchmark is never the bottleneck.
@bindureddy Hey Bindu, been following your work on Abacus and would love to write on it for our newsletter, feel free to send me a message so I can explain more!
@bindureddy They abandoned Sonnet
@bindureddy multi-turn codegen is where it gets ugly the token guzzle compounds. promo pricing makes turn 1 look clean, by turn 4 you're paying opus 4.8 prices for a worse result
@bindureddy Couldn't agree more. I used it for ONE prompt and immediately returned to Sonnet 4.6. What used to be 1-3% session usage is consuming 12-18% and the results don't warrant it. That's closer to Opus token usage.
@bindureddy Will stay connected for further update, need to know how actually it works
@bindureddy Cost-per-quality is the right framing and promotional pricing muddies the comparison — the real test is August when prices normalize. Would be curious what you see on LiveBench once the token guzzling is priced at full rate vs Opus 4.8.
@bindureddy Yeah, probably a teaser. Access termination coming in couple of days 🤣
@bindureddy “Promo price helps” — meaning your conclusion is built on a temporary discount. What happens to your verdict when pricing goes back to standard?
@bindureddy Benchmark gap matters less than which one self-corrects in a loop. Gave an agent one prompt — it built a bilingual invoice app, caught its own PDF bug, and fixed it. That "tests its own output" loop beats a LiveBench point. Multi-turn agentic reliability is the real test.