Snaki (@Snak1i) header
Snaki (@Snak1i) profile picture

Trying to figure out crypto Building my YouTube journey

Ukraine
Joined June 2025
Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)Photo by Snaki (@Snak1i)
a 15-minute up/down on gold or oil fits a trading agent better than a person
Commodities, up or down? Gold, silver, oil, & natural gas up/downs are now live on Polymarket. 15 minute, hourly, & daily.
2
3
Commodities, up or down? Gold, silver, oil, & natural gas up/downs are now live on Polymarket. 15 minute, hourly, & daily.
29
28
18
97
94,746
paweł huryn's test: 2 real repos, 105 replanted bugs, only fixed bugs count. score and cost: - gpt-6.2 astra (max): 62.7 for $28.35 - fable 5.2 (max): 58 for $139.65 - sonnet 5.5 (max): 51.3 for $154.03 sonnet 5.5 lists at a quarter of fable's price for input and output ($2/$10 vs $8/$40). it still cost 10% more here, so at max effort it went through several times fable's tokens. dollars per point: astra $0.45, fable $2.41, sonnet $3.00. before you default to max on a cheaper model, compare effort levels on your own tasks.
I tested GPT-6.2 Astra on real work. 2 repos, 105 replanted bugs. Find and fix what you can. It's the new #1. The results: - GPT-6.2 Astra (max): 62.7 for $28.35 - GPT-6.2 Astra (xhigh): 60 for $23.53 - Fable 5.2 (max): 58 for $139.65 - Gemini 4 Carbon (high): 57.3 for $31.32 - Sonnet 5.5 (max): 51.3 for $154.03 - GPT-6 Astra (max): 45 for $33.03 From GPT-6 Astra to GPT-6.2 Astra: +17.7 points at max, +17 at xhigh. Whatever people say about other use cases, it's a code review beast. GPT-6.2 Sol and more effort levels drop soon in the thread 🧵
2
6
anthropic added an experimental code mode to claude code yesterday. claude writes javascript that makes the tool calls, and the code can run commands in parallel, pass one result into the next call, and return only the output claude needs to read. turn it on: CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude what to know first: - interactive terminal only. not -p, the agent sdk, vs code or the desktop app - v2.1.323+, with opus 5.5, sonnet 5.5, haiku 5.5 or fable 5.2 - project settings ignore the variable, so a cloned repo can't switch it on - a code mode session resumes only with code mode on deny and ask rules still check every call inside the script. in auto mode the classifier reviews the code before it runs, not each call.
1
1
10
opus 5.5, a 100k-token prompt prefix, one call every 10 minutes for an hour (6 calls): - no cache: $2.40 - 5-minute cache: $3.00. it expires between calls, so every call pays the 1.25x write again - 1-hour cache: $0.90. one 2x write, then five hits at $0.20 per million when calls land more than 5 minutes apart, the default cache costs more than no cache at all. for agents that wait on people or slow tools, set "ttl": "1h" in cache_control.
1
1
11
the shortest prompt claude will cache, per anthropic's docs: - 512 tokens: fable 5.2, opus 5.5, opus 5, sonnet 5.5, haiku 5.5 - 1,024: sonnet 5, sonnet 4.6 - 4,096: haiku 4.5 anything shorter runs without caching, even with cache_control set, and no error comes back. so a 900-token system prompt caches on sonnet 5.5 and quietly doesn't on sonnet 5. check the usage block: if cache_creation_input_tokens and cache_read_input_tokens are both 0, nothing was cached.
1
2
12
sonnet 4.6 → sonnet 5.5 cuts the price per token by a third, $3/$15 → $2/$10. two lines in anthropic's migration guide eat into that: - the same text produces about 30% more tokens on the new tokenizer - a 2000×1500 image costs about 2.5x as many tokens (high-res tier, up to 4,784 tokens per image) net, before caching: text comes out about 13% cheaper, that image about 67% more expensive. recount your prompts with the token counting endpoint before you budget the switch.
1
1
10
two things your sonnet 5 code may send that sonnet 5.5 rejects with a 400, per anthropic's migration guide: - thinking: {type: "disabled"}. send {type: "between_tools"} instead, at low, medium or high effort only - tool_choice "any" or "tool". send auto and mark the tool strict: true effort still defaults to high, but the levels are recalibrated. the guide says to start at medium for well-specified tasks. the price stays the same as sonnet 5, except cache reads: $0.10 per million, half the old rate. test medium on your own tasks before you keep high as the default.
1
1
15
deepnote's agent benchmark counted tool calls per task (score out of 10): - sonnet 5.5: 8.9 calls, 6.65 - opus 5.5: 10.0 calls, 7.84 - fable 5.2: 20.9 calls, 8.17 - gemini 4 argon: 34.5 calls, 5.60 - grok 4.7: 35.5 calls, 4.51 up to about 20 calls, the score climbed in these rows. argon and grok, at 35, scored lowest of all 12 models and timed out before a useful answer on 33% of tasks. if your agent averages 30+ tool calls a task, log where the calls go before you add more tools.
1
2
17
ai prices worth keeping from this week: - gemini 4 hydrogen launched at $1.50 in / $7.50 out per million tokens - gemini 4 argon and carbon: $2 / $10 through nov 30, $4 / $20 from dec 1 - claude fable 5.2: $8 / $40 (fable 5.1 was $10 / $50) - claude opus 5.5 cache hit: $0.20 per million (opus 5 was $0.50) - deepnote's agent benchmark: fable 5.2 at $1.87 a task, gpt-6.1 sol at $0.11 argon and carbon double on december 1. if they're in your stack, price december now.
2
1
13
the latest claude releases cut cache hit prices harder than list prices on opus and sonnet. per million tokens: - opus 5 → 5.5: input $5 → $4, cache hit $0.50 → $0.20 (-60%) - sonnet 5 → 5.5: input $2 → $2, cache hit $0.20 → $0.10 (-50%) - fable 5.1 → 5.2: input $10 → $8, cache hit $0.25 → $0.20 (-20%) an agent that rereads a 200k context 50 times pays $5.00 for those reads on opus 5 and $2.00 on opus 5.5. check usage.cache_read_input_tokens on your agent calls. if it's zero, none of this cut reaches you.
1
1
17
anthropic's migration guide lists three things that change when you swap the model id from opus 5 to opus 5.5: - effort now defaults to medium. opus 5 defaulted to high - thinking: {type: "disabled"} returns a 400. remove the thinking field and pick an effort level - tool_choice "any" or "tool" returns a 400. use auto with strict tool use, and say in the prompt when the tool applies the price drops from $5/$25 to $4/$20 per million tokens. set effort explicitly and rerun your evals before you ship the swap.
1
1
12