Model guidance

Rule: An upgrade pays when it adds at least 5 points (or 75 Elo on GDPval) for each doubling of cost per task. Below that, the cheaper setting is the sweet spot.

Latest news

2026-10-15 · Argon introductory API pricing — $2 input / $10 output per million tokens through November 30, then $4 / $20. API offer only; subscription allowances are separate.
2026-10-13 · Gemini 4 Argon broadly available — Independent AutomationBench: 78% versus Sol Medium 63%, at $1.99 versus $0.21 per benchmark task. More capability comes at a higher task cost. Benchmarks
2026-10-07 · Sonnet 5.5 cache reads halved — Cache reads fell to $0.10 per million tokens. Anthropic estimates about 20% lower cost on most agentic tasks; this is a vendor claim, not a measured subscription saving.

ChatGPT and Codex

Mechanical text work | GPT-6 Luna max ($0.07, 38), best when the step calls tools (AutomationBench 53%) | GPT-6.1 Sol low ($0.13, 42)

Scripted relays and tool workflows | GPT-6.1 Sol medium ($0.21, 63%) | Rarely. GPT-6.1 Sol xhigh ($0.39, 67%) adds 4 points

Documents from supplied facts | GPT-6 Luna max (1432, $0.07) | GPT-6.1 Sol max (1594, $0.72)

People, messages, money, records | GPT-6.1 Sol high ($0.32, 41), verify names and figures | Rarely. Sol max ($0.72, 42) adds 1 point

Computer use and browser work | GPT-6 Luna max (49%, $0.22). Sol and Astra not yet evaluated on this chart | none

Agentic coding | GPT-6.1 Sol xhigh ($0.39, 54%) | Rarely. GPT-6.1 Sol max ($0.72, 56%) adds 2 points

Small, clear coding fixes | GPT-6.1 Sol medium ($0.21, 48%) | GPT-6.1 Sol xhigh ($0.39, 54%)

Research and reasoning | GPT-6.1 Sol medium ($0.21, 50%) | GPT-6 Astra max ($3.26, 55%), rarely worth it

General agentic judgment | GPT-6.1 Sol medium ($0.21, 48) | GPT-6.1 Sol high ($0.32, 50)

Claude

Mechanical text work | Haiku 5.5 low ($0.02, 29) | Haiku 5.5 medium ($0.05, 34) if low misses

Scripted relays and tool workflows | Sonnet 5.5 medium ($0.48, 55%). Opus 5.5 low ($0.55, 53%) when the run touches names or facts | Opus 5.5 medium ($1.34, 61%) when the workflow branches

Documents from supplied facts | Haiku 5.5 max (pilot): 1618 at $0.21, near Sonnet 5.5 xhigh at a tenth of the cost | Rarely. Sonnet 5.5 xhigh (1728, $2.01) adds 110 Elo for 9.6x the cost

People, messages, money, records | Opus 5.5 low ($0.55, 39) | Rarely. Opus 5.5 max adds 7 points for 11x the cost

Computer use and browser work | Haiku 5.5 xhigh (pilot, 67.5%, $0.30) | Haiku 5.5 max (72.4%, $0.62) ties Sonnet 5.5 high (73%, $1.50). Sonnet 5.5 xhigh (81%, $2.30) for must-succeed runs

Agentic coding | Opus 5.5 medium ($1.34, 53%) | Opus 5.5 high ($1.82, 57%) for hard bugs

Small, clear coding fixes | Haiku 5.5 max (pilot, $0.21, 33%), ahead of Opus 5.5 low (31%, $0.55) | Opus 5.5 medium if the fix stalls

Research and reasoning | Opus 5.5 low ($0.55, 48%). Vendor chart shows Sonnet 5.5 high at 48% for $0.08 per question | Opus 5.5 high ($1.82, 56%)

General agentic judgment | Opus 5.5 medium ($1.34, 51) | Opus 5.5 high ($1.82, 54)

Gemini/Antigravity

Mechanical text work | Gemini 3.8 Flash low (index 33, cost per task not published), also for deterministic scripts | Medium ($0.93, 40)

Scripted relays and tool workflows | Gemini 3.8 Flash medium ($0.93, 61%) | None. High scores 60%

Documents from supplied facts | Gemini 3.8 Flash medium (1421, $0.93) | High adds 15 Elo, not worth it

People, messages, money, records | Gemini 3.8 Flash medium ($0.93, 29), below Opus 5.5 low and Sol low | none

Agentic coding | Gemini 3.8 Flash medium ($0.93, 20%) for agent loops, refactors, and hard bugs. Weak for coding | High also scores 20%

Research and reasoning | Gemini 3.8 Flash high ($1.24, 48%) | None. High is the pick here (HLE 48% vs 42% at medium)

General agentic judgment | Gemini 3.8 Flash medium ($0.93, 40) | High adds 1 point, not worth it

Grok

Documents from supplied facts | Grok 4.6 medium (1612, $1.14) | Grok 4.7 high (1706, $2.73)

People, messages, money, records | Grok 4.6 low ($0.38, 26), verify names and figures | Rarely. Grok 4.7 xhigh ($3.74, 32) adds 6 points

Agentic coding | Grok 4.6 high ($1.48, 21%) | Rarely. Grok 4.7 high ($2.73, 25%) adds 4 points

Small, clear coding fixes | Grok 4.6 high ($1.48, 21%) | Rarely. Grok 4.7 high ($2.73, 25%) adds 4 points

Research and reasoning | Grok 4.6 medium ($1.14, 42%) | Rarely. Grok 4.6 xhigh ($1.75, 44%) adds 2 points

General agentic judgment | Grok 4.6 medium ($1.14, 43) | Rarely. Grok 4.7 high ($2.73, 46) adds 3 points