{
 "updated": "2026-10-15",
 "revision": "5e991ea972cbba3ff5772dd559d754964812d2abf14d41ad733eaba1e63ae58a",
 "news": [
  {
   "id": "argon-api-intro",
   "vendor": "Google",
   "date": "2026-10-15",
   "title": "Argon introductory API pricing",
   "summary": "$2 input / $10 output per million tokens through November 30, then $4 / $20. API offer only; subscription allowances are separate.",
   "source": "https://ai.google.dev/gemini-api/docs/pricing?hl=en",
   "checked": "2026-10-15",
   "expires_at": "2026-12-01T00:00:00Z"
  },
  {
   "id": "argon-release",
   "vendor": "Google",
   "date": "2026-10-13",
   "title": "Gemini 4 Argon broadly available",
   "summary": "Independent AutomationBench: 78% versus Sol Medium 63%, at $1.99 versus $0.21 per benchmark task. More capability comes at a higher task cost.",
   "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/",
   "checked": "2026-10-15",
   "expires_at": "2026-11-12T00:00:00Z",
   "independent_source": "https://artificialanalysis.ai/models/comparisons/gemini-4-argon-vs-gpt-6-1-sol-medium"
  },
  {
   "id": "sonnet-cache-cut",
   "vendor": "Anthropic",
   "date": "2026-10-07",
   "title": "Sonnet 5.5 cache reads halved",
   "summary": "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.",
   "source": "https://www.anthropic.com/claude-haiku-5-5",
   "checked": "2026-10-15",
   "expires_at": "2026-11-06T00:00:00Z"
  }
 ],
 "review": {
  "checked": "2026-10-15",
  "verified_points": 47,
  "pending_points": [
   "Luna none: attempted comparison returned HTTP 404",
   "Grok 4.7 medium: attempted comparison returned HTTP 404",
   "Vendor chart images: approximate points retained from 2026-10-08",
   "OpenAI Pro 200 temporary allowance: official terms confirm eligible subscriptions active September 22 through September 29 at 10 AM Pacific keep the previous allowance through October 29. Exact promotion start and expiry timezone, specific account eligibility and any credit grant are unverified. No API discount or credit amount assumed. Source checked 2026-10-15: https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers",
   "OpenAI referral rewards vary by account and offer. Banked Work/Codex resets normally expire 30 days after award; exact account offer, caps and award date are unverified. Not API credits. No price adjustment. Source checked 2026-10-15: https://help.openai.com/en/articles/20001271-chatgpt-desktop-referral-promotions",
   "Google Antigravity promotional overage credits require account-specific terms. No verified temporary subscription boost found in reviewed official plans/pricing. No API discount inferred. Sources checked 2026-10-15: https://antigravity.google/docs/plans and https://antigravity.google/pricing",
   "Grok for Startups may cover API and other eligible services only after approval. Amount and expiry come from approval terms, normally six months after credit application; account eligibility and grant are unverified. Cannot stack without written approval. No token-price adjustment. Source checked 2026-10-15: https://x.ai/legal/startup-program-terms",
   "Gemini 4 Argon low/medium: attempted comparison URLs returned HTTP 404; no scores inferred."
  ],
  "note": "All 47 tracked independent model/effort points verified from comparison tables on 2026-10-15. Added 14 previously missing effort points. GDPval Elo can move as comparison results update. Vendor chart points remain approximate and dated 2026-10-08. Unpublished costs and benchmark scores remain null. Independent reasoning scores do not describe the local non-thinking QAT instance. Overall and task-specific model picks retained after review. Account-specific offers without verified eligibility and complete dates are excluded from price calculations. The newly added model has one verified effort point; other attempted effort sources remain pending. Existing task-model picks remain in place pending the new model access and effort comparison review; news and charts do not assume account eligibility."
 },
 "overall_value_leader": {
  "model": "GPT-6.1 Sol",
  "effort": "medium",
  "reviewed": "2026-10-15",
  "basis": "Balanced capability and cost for everyday, multistep work. Use the task-specific picks below when your work needs something different.",
  "source": "https://artificialanalysis.ai/models/gpt-6-1-sol-medium",
  "task_id": "judgment"
 },
 "promotions": [
  {
   "id": "google-gemini38-intro-2026",
   "vendor": "Google",
   "model": "Gemini 3.8 Flash",
   "scope": "Google AI Studio / Gemini Developer API paid Standard tier; not a subscription-plan discount",
   "starts_at": null,
   "ends_at": "2027-01-01T00:00:00Z",
   "boundary_note": "Conservative UTC cutoff: vendor lists dates but does not specify a timezone. Use January 1 at 00:00 UTC as a conservative display cutoff; billing follows vendor terms.",
   "promotional_multiplier": 0.5,
   "promotional_rates": {
    "input": 0.75,
    "output": 3.75,
    "cached_input": 0.075
   },
   "standard_rates": {
    "input": 1.5,
    "output": 7.5,
    "cached_input": 0.15
   },
   "source": "https://ai.google.dev/gemini-api/docs/pricing",
   "checked": "2026-10-15",
   "task_estimate_note": "AA reported medium/high costs already use introductory input/output/cache rates. Standard task estimates assume identical token usage; not a fresh benchmark run.",
   "discount_label": "50% introductory API pricing",
   "last_discounted_date": "2026-12-31",
   "standard_from": "2027-01-01"
  },
  {
   "id": "google-argon-intro-2026",
   "vendor": "Google",
   "model": "Gemini 4 Argon",
   "scope": "Gemini Developer API paid Standard tier. No free tier. Subscription eligibility and Antigravity allowances are separate.",
   "starts_at": null,
   "ends_at": "2026-12-01T00:00:00Z",
   "boundary_note": "Conservative UTC display cutoff because the vendor states December 1 without a timezone; vendor billing terms remain authoritative. Exact offer start timestamp is not specified.",
   "promotional_multiplier": 0.5,
   "promotional_rates": {
    "input": 2,
    "output": 10,
    "cached_input": 0.1
   },
   "standard_rates": {
    "input": 4,
    "output": 20,
    "cached_input": 0.2
   },
   "source": "https://ai.google.dev/gemini-api/docs/pricing?hl=en",
   "checked": "2026-10-15",
   "task_estimate_note": "AA high-effort cost $1.99 already includes introductory rates; post-expiry $3.98 assumes unchanged token usage and is an estimate, not a fresh benchmark measurement.",
   "discount_label": "50% introductory API pricing",
   "last_discounted_date": "2026-11-30",
   "standard_from": "2026-12-01"
  }
 ],
 "subscription_promotions": [
  {
   "id": "claude-artifacts-oct2026",
   "vendor": "Anthropic",
   "name": "Claude artifact usage promotion",
   "starts_at": "2026-10-01T18:00:00Z",
   "ends_at": "2026-10-16T07:00:00Z",
   "checked": "2026-10-15",
   "api_applies": false,
   "weekly_limit_applies": false,
   "scope": "Free (from October 7), Pro, Max and Team; excludes Enterprise. Eligible Claude chats and cloud tasks in the Claude app only.",
   "benefit": "50% less five-hour session usage for eligible artifact work; approximately 2x eligible session capacity, not a blanket token discount.",
   "limits": "Chat: next 10 messages after an artifact edit, up to the first 15 steps per reply. Cloud tasks in the Claude app: about 45 minutes when starting with Document, Presentation or Design, or up to 80 steps after an artifact edit.",
   "exclusions": "No Claude API, Claude Code, local desktop tasks in the Claude app, scheduled/unattended tasks, usage-credit charges, older artifacts, standalone Claude Design, or Google Drive-created Docs/Sheets/Slides. Weekly allowance is unchanged.",
   "expires_label": "October 15, 2026 at 11:59 PM Pacific",
   "recommendation": "For eligible document, slide and design work, consider Claude during this offer to stretch the five-hour allowance. Choose the lowest model and effort that passes your checks. API benchmark costs do not measure subscription allowance; no cross-vendor subscription winner is established.",
   "source": "https://support.claude.com/en/articles/17274727-artifact-usage-promotion"
  }
 ],
 "source": "https://robbagley-afk.github.io/model-value-map/",
 "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.",
 "platforms": {
  "claude": [
   {
    "task": "Mechanical text work",
    "pick": "Haiku 5.5 low ($0.02, 29)",
    "upgrade": "Haiku 5.5 medium ($0.05, 34) if low misses"
   },
   {
    "task": "Scripted relays and tool workflows",
    "pick": "Sonnet 5.5 medium ($0.48, 55%). Opus 5.5 low ($0.55, 53%) when the run touches names or facts",
    "upgrade": "Opus 5.5 medium ($1.34, 61%) when the workflow branches",
    "budget": "Sonnet 5.5 low ($0.35, 49%)",
    "callouts": [
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol medium scores 63% vs 55% at $0.21 per task, 44% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Documents from supplied facts",
    "pick": "Haiku 5.5 max (pilot): 1618 at $0.21, near Sonnet 5.5 xhigh at a tenth of the cost",
    "upgrade": "Rarely. Sonnet 5.5 xhigh (1728, $2.01) adds 110 Elo for 9.6x the cost",
    "budget": "Haiku 5.5 xhigh ($0.12, 1513)"
   },
   {
    "task": "People, messages, money, records",
    "pick": "Opus 5.5 low ($0.55, 39)",
    "upgrade": "Rarely. Opus 5.5 max adds 7 points for 11x the cost",
    "callouts": [
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol medium scores 40 vs 39 at $0.21 per task, 38% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Computer use and browser work",
    "pick": "Haiku 5.5 xhigh (pilot, 67.5%, $0.30)",
    "upgrade": "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",
    "mid": "Haiku 5.5 medium ($0.13, 53.5%)",
    "budget": "Haiku 5.5 low ($0.07, 42%)"
   },
   {
    "task": "Agentic coding",
    "pick": "Opus 5.5 medium ($1.34, 53%)",
    "upgrade": "Opus 5.5 high ($1.82, 57%) for hard bugs",
    "mid": "Sonnet 5.5 high ($0.88, 44%)",
    "budget": "Haiku 5.5 max ($0.21, 33%)",
    "exception": "Opus 5.5 high stays the upgrade for hard bugs; medium covers most coding at 74% of the cost",
    "callouts": [
     "Rule exception: Opus 5.5 high stays the upgrade for hard bugs; medium covers most coding at 74% of the cost",
     "Smaller model beats larger: Haiku 5.5 max ($0.21, 33%) scores at least as well as Opus 5.5 low ($0.55, 31%).",
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol xhigh scores 54% vs 53% at $0.39 per task, 29% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Small, clear coding fixes",
    "pick": "Haiku 5.5 max (pilot, $0.21, 33%), ahead of Opus 5.5 low (31%, $0.55)",
    "upgrade": "Opus 5.5 medium if the fix stalls",
    "budget": "Haiku 5.5 xhigh ($0.12, 29%)",
    "exception": "A deliberately cheap first try for small fixes; move to Opus 5.5 medium if the fix stalls",
    "callouts": [
     "Rule exception: A deliberately cheap first try for small fixes; move to Opus 5.5 medium if the fix stalls"
    ]
   },
   {
    "task": "Research and reasoning",
    "pick": "Opus 5.5 low ($0.55, 48%). Vendor chart shows Sonnet 5.5 high at 48% for $0.08 per question",
    "upgrade": "Opus 5.5 high ($1.82, 56%)",
    "budget": "Haiku 5.5 max ($0.21, 44%)",
    "exception": "Haiku 5.5 max scores close but hallucinates often (AA-Omniscience 11 vs 39), so Opus 5.5 low stays the floor; step up for hard problems",
    "callouts": [
     "Rule exception: Haiku 5.5 max scores close but hallucinates often (AA-Omniscience 11 vs 39), so Opus 5.5 low stays the floor; step up for hard problems",
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol medium scores 50% vs 48% at $0.21 per task, 38% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "General agentic judgment",
    "pick": "Opus 5.5 medium ($1.34, 51)",
    "upgrade": "Opus 5.5 high ($1.82, 54)",
    "budget": "Haiku 5.5 max ($0.21, 43)",
    "exception": "Opus 5.5 medium is the maintained general-purpose pick; Haiku 5.5 max trails on fact reliability (AA-Omniscience 11 vs 40), and high stays the upgrade",
    "callouts": [
     "Rule exception: Opus 5.5 medium is the maintained general-purpose pick; Haiku 5.5 max trails on fact reliability (AA-Omniscience 11 vs 40), and high stays the upgrade",
     "Smaller model beats larger: Haiku 5.5 max ($0.21, 43) scores at least as well as Opus 5.5 low ($0.55, 42).",
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol xhigh scores 51 vs 51 at $0.39 per task, 29% of the cost. API cost, not subscription usage."
    ]
   }
  ],
  "codex": [
   {
    "task": "Mechanical text work",
    "pick": "GPT-6 Luna max ($0.07, 38), best when the step calls tools (AutomationBench 53%)",
    "upgrade": "GPT-6.1 Sol low ($0.13, 42)",
    "budget": "GPT-6 Luna xhigh ($0.04, 35)",
    "exception": "Luna max is the pick when the step calls tools (AutomationBench 53% vs 48% at xhigh)",
    "callouts": [
     "Rule exception: Luna max is the pick when the step calls tools (AutomationBench 53% vs 48% at xhigh)"
    ]
   },
   {
    "task": "Scripted relays and tool workflows",
    "pick": "GPT-6.1 Sol medium ($0.21, 63%)",
    "upgrade": "Rarely. GPT-6.1 Sol xhigh ($0.39, 67%) adds 4 points",
    "budget": "GPT-6 Luna max ($0.07, 53%)",
    "callouts": [
     "Smaller model beats larger: GPT-6 Luna max ($0.07, 53%) scores at least as well as GPT-6.1 Sol low ($0.13, 53%)."
    ]
   },
   {
    "task": "Documents from supplied facts",
    "pick": "GPT-6 Luna max (1432, $0.07)",
    "upgrade": "GPT-6.1 Sol max (1594, $0.72)",
    "budget": "GPT-6 Luna xhigh ($0.04, 1364)"
   },
   {
    "task": "People, messages, money, records",
    "pick": "GPT-6.1 Sol high ($0.32, 41), verify names and figures",
    "upgrade": "Rarely. Sol max ($0.72, 42) adds 1 point",
    "budget": "GPT-6.1 Sol low ($0.13, 38)",
    "exception": "Sol high stays for accuracy on names and figures, where one wrong fact costs more than the extra $0.11 per task",
    "callouts": [
     "Rule exception: Sol high stays for accuracy on names and figures, where one wrong fact costs more than the extra $0.11 per task"
    ]
   },
   {
    "task": "Computer use and browser work",
    "pick": "GPT-6 Luna max (49%, $0.22). Sol and Astra not yet evaluated on this chart",
    "budget": "GPT-6 Luna medium ($0.13, 37%)",
    "callouts": [
     "Cheaper elsewhere: Anthropic Haiku 5.5 medium scores 53.5% vs 49% at $0.13 per task, 59% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Agentic coding",
    "pick": "GPT-6.1 Sol xhigh ($0.39, 54%)",
    "upgrade": "Rarely. GPT-6.1 Sol max ($0.72, 56%) adds 2 points",
    "budget": "GPT-6.1 Sol medium ($0.21, 48%)"
   },
   {
    "task": "Small, clear coding fixes",
    "pick": "GPT-6.1 Sol medium ($0.21, 48%)",
    "upgrade": "GPT-6.1 Sol xhigh ($0.39, 54%)",
    "budget": "GPT-6.1 Sol low ($0.13, 31%)",
    "exception": "A cheap first try for small fixes, matching the Claude row; move to Sol xhigh, the agentic coding pick, if the fix stalls",
    "callouts": [
     "Rule exception: A cheap first try for small fixes, matching the Claude row; move to Sol xhigh, the agentic coding pick, if the fix stalls"
    ]
   },
   {
    "task": "Research and reasoning",
    "pick": "GPT-6.1 Sol medium ($0.21, 50%)",
    "upgrade": "GPT-6 Astra max ($3.26, 55%), rarely worth it",
    "budget": "GPT-6.1 Sol low ($0.13, 47%)",
    "exception": "Sol medium adds 3 points over low for $0.08 more per task; kept for multi-step reasoning, where a miss costs more than the difference",
    "callouts": [
     "Rule exception: Sol medium adds 3 points over low for $0.08 more per task; kept for multi-step reasoning, where a miss costs more than the difference"
    ]
   },
   {
    "task": "General agentic judgment",
    "pick": "GPT-6.1 Sol medium ($0.21, 48)",
    "upgrade": "GPT-6.1 Sol high ($0.32, 50)",
    "budget": "GPT-6.1 Sol low ($0.13, 42)"
   }
  ],
  "antigravity": [
   {
    "task": "Mechanical text work",
    "pick": "Gemini 3.8 Flash low (index 33, cost per task not published), also for deterministic scripts",
    "upgrade": "Medium ($0.93, 40)",
    "callouts": [
     "New model not yet placed: Gemini 4 Argon high ($1.99, 53) is tracked but not yet reviewed for this row."
    ]
   },
   {
    "task": "Scripted relays and tool workflows",
    "pick": "Gemini 3.8 Flash medium ($0.93, 61%)",
    "upgrade": "None. High scores 60%",
    "callouts": [
     "New model not yet placed: Gemini 4 Argon high ($1.99, 78%) is tracked but not yet reviewed for this row.",
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol medium scores 63% vs 61% at $0.21 per task, 23% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Documents from supplied facts",
    "pick": "Gemini 3.8 Flash medium (1421, $0.93)",
    "upgrade": "High adds 15 Elo, not worth it",
    "callouts": [
     "New model not yet placed: Gemini 4 Argon high ($1.99, 1624) is tracked but not yet reviewed for this row.",
     "Cheaper elsewhere: OpenAI GPT-6 Luna max scores 1432 vs 1421 at $0.07 per task, 8% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "People, messages, money, records",
    "pick": "Gemini 3.8 Flash medium ($0.93, 29), below Opus 5.5 low and Sol low",
    "callouts": [
     "New model not yet placed: Gemini 4 Argon high ($1.99, 42) is tracked but not yet reviewed for this row.",
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol low scores 38 vs 29 at $0.13 per task, 14% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Computer use and browser work",
    "pick": "Not yet evaluated"
   },
   {
    "task": "Agentic coding",
    "pick": "Gemini 3.8 Flash medium ($0.93, 20%) for agent loops, refactors, and hard bugs. Weak for coding",
    "upgrade": "High also scores 20%",
    "callouts": [
     "New model not yet placed: Gemini 4 Argon high ($1.99, 57%) is tracked but not yet reviewed for this row.",
     "Cheaper elsewhere: Anthropic Haiku 5.5 high scores 22% vs 20% at $0.08 per task, 9% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Small, clear coding fixes",
    "pick": "Gemini 3.8 Flash medium ($0.93, 20%)",
    "callouts": [
     "New model not yet placed: Gemini 4 Argon high ($1.99, 57%) is tracked but not yet reviewed for this row.",
     "Cheaper elsewhere: Anthropic Haiku 5.5 high scores 22% vs 20% at $0.08 per task, 9% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Research and reasoning",
    "pick": "Gemini 3.8 Flash high ($1.24, 48%)",
    "upgrade": "None. High is the pick here (HLE 48% vs 42% at medium)",
    "budget": "Gemini 3.8 Flash medium ($0.93, 42%)",
    "callouts": [
     "New model not yet placed: Gemini 4 Argon high ($1.99, 57%) is tracked but not yet reviewed for this row.",
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol medium scores 50% vs 48% at $0.21 per task, 17% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "General agentic judgment",
    "pick": "Gemini 3.8 Flash medium ($0.93, 40)",
    "upgrade": "High adds 1 point, not worth it",
    "callouts": [
     "New model not yet placed: Gemini 4 Argon high ($1.99, 53) is tracked but not yet reviewed for this row.",
     "Cheaper elsewhere: Anthropic Haiku 5.5 xhigh scores 41 vs 40 at $0.12 per task, 13% of the cost. API cost, not subscription usage."
    ]
   }
  ],
  "grok": [
   {
    "task": "Mechanical text work",
    "pick": "Grok 4.6 low ($0.38, 35)",
    "upgrade": "Grok 4.6 medium ($1.14, 43)",
    "callouts": [
     "Cheaper elsewhere: OpenAI GPT-6 Luna xhigh scores 35 vs 35 at $0.04 per task, 11% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Scripted relays and tool workflows",
    "pick": "Grok 4.6 high ($1.48, 67%)",
    "budget": "Grok 4.6 low ($0.38, 53%)",
    "callouts": [
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol xhigh scores 67% vs 67% at $0.39 per task, 26% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Documents from supplied facts",
    "pick": "Grok 4.6 medium (1612, $1.14)",
    "upgrade": "Grok 4.7 high (1706, $2.73)",
    "budget": "Grok 4.6 low ($0.38, 1410)",
    "callouts": [
     "Cheaper elsewhere: Anthropic Haiku 5.5 max scores 1618 vs 1612 at $0.21 per task, 18% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "People, messages, money, records",
    "pick": "Grok 4.6 low ($0.38, 26), verify names and figures",
    "upgrade": "Rarely. Grok 4.7 xhigh ($3.74, 32) adds 6 points",
    "callouts": [
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol low scores 38 vs 26 at $0.13 per task, 34% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Computer use and browser work",
    "pick": "Not yet evaluated"
   },
   {
    "task": "Agentic coding",
    "pick": "Grok 4.6 high ($1.48, 21%)",
    "upgrade": "Rarely. Grok 4.7 high ($2.73, 25%) adds 4 points",
    "callouts": [
     "Cheaper elsewhere: Anthropic Haiku 5.5 high scores 22% vs 21% at $0.08 per task, 5% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Small, clear coding fixes",
    "pick": "Grok 4.6 high ($1.48, 21%)",
    "upgrade": "Rarely. Grok 4.7 high ($2.73, 25%) adds 4 points",
    "callouts": [
     "Cheaper elsewhere: Anthropic Haiku 5.5 high scores 22% vs 21% at $0.08 per task, 5% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "Research and reasoning",
    "pick": "Grok 4.6 medium ($1.14, 42%)",
    "upgrade": "Rarely. Grok 4.6 xhigh ($1.75, 44%) adds 2 points",
    "callouts": [
     "Cheaper elsewhere: Anthropic Haiku 5.5 xhigh scores 43% vs 42% at $0.12 per task, 11% of the cost. API cost, not subscription usage."
    ]
   },
   {
    "task": "General agentic judgment",
    "pick": "Grok 4.6 medium ($1.14, 43)",
    "upgrade": "Rarely. Grok 4.7 high ($2.73, 46) adds 3 points",
    "budget": "Grok 4.6 low ($0.38, 35)",
    "callouts": [
     "Cheaper elsewhere: OpenAI GPT-6.1 Sol medium scores 48 vs 43 at $0.21 per task, 18% of the cost. API cost, not subscription usage."
    ]
   }
  ],
  "local": [
   {
    "task": "Mechanical text work",
    "pick": "Gemma 4 26B-A4B ($0). In-house test: classification 6/6, reformatting 8/8, instruction following 6/6 on all three runs",
    "upgrade": "Use Haiku 5.5 low when output must be valid JSON. Gemma broke JSON on 1 to 2 of 6 extractions after audit"
   },
   {
    "task": "Scripted relays and tool workflows",
    "pick": "Light use only. Tool calls 3/4 in-house, AutomationBench 2% (AA estimate). Validate every call",
    "upgrade": "Sonnet 5.5 low"
   },
   {
    "task": "Documents from supplied facts",
    "pick": "Reformatting only. GDPval 566 (AA estimate)"
   },
   {
    "task": "People, messages, money, records",
    "pick": "Never for recall. AA-Omniscience -51. It did answer UNKNOWN to 8/8 made-up facts when offered that option"
   },
   {
    "task": "Computer use and browser work",
    "pick": "Not yet evaluated"
   },
   {
    "task": "Agentic coding",
    "pick": "Gemma: Terminal-Bench 2.0, our run, 4-bit, thinking off: 15% (13 of 89). Older test, not on this chart"
   },
   {
    "task": "Small, clear coding fixes",
    "pick": "Not yet evaluated"
   },
   {
    "task": "Research and reasoning",
    "pick": "Not suitable. HLE 8.7% (Google card), money math 4/6 in-house"
   },
   {
    "task": "General agentic judgment",
    "pick": "Not suitable (index 17)"
   }
  ]
 },
 "best_value": [
  {
   "task": "Mechanical text work",
   "best": "Gemma on a local LM Studio server ($0 cloud). In the cloud, Haiku 5.5 low ($0.02, index 29)",
   "avoid": "Sonnet or Opus for pure mechanical work. Gemma for anything that calls tools (AutomationBench 2%). Grok for mechanical work: Grok 4.6 low ($0.38, 35) costs 19x Haiku 5.5 low for 6 more points. Gemini 3.8 Flash medium ($0.93, 40) costs 13x GPT-6 Luna max for 2 more points. With Gemini, use Flash low (33) for deterministic scripts Newly measured Luna low ($0.0045, index 22) is a candidate for narrow supplied-input steps after a task check, not an established replacement for the maintained picks."
  },
  {
   "task": "Scripted relays and tool workflows",
   "best": "GPT-6 Luna max ($0.07, AutomationBench 53%), which matches Opus 5.5 low at 13% of the cost. Argon high (78%, $1.99 introductory through November 30) is the highest tracked AutomationBench score under $3; eligibility and task fit still need checking",
   "avoid": "Haiku 5.5 at any effort (36% at xhigh). Opus 5.5 high and up, Sonnet 5.5 xhigh and max: 1 to 2 points per cost doubling. Gemini 3.8 Flash high (60%) scores below medium ($0.93, 61%), which already trails GPT-6.1 Sol medium (63%, $0.21) Argon high costs $1.99 per AA task through November 30; estimated $3.98 from December 1 assuming unchanged tokens. Subscription caps and availability are separate."
  },
  {
   "task": "Documents from supplied facts",
   "best": "Haiku 5.5 max (GDPval 1618, $0.21). GPT-6 Luna max (1432, $0.07) if cost matters more than polish",
   "avoid": "Opus 5.5 high (1704) adds 86 Elo for 8.7x the cost. Sonnet 5.5 max. Haiku hallucinates often (AA-Omniscience 6 to 11), so supply the facts and check names and numbers. Grok 4.6 medium (1612, $1.14) roughly ties Haiku 5.5 max (1618, $0.21) at 5x the cost. Gemini 3.8 Flash medium (1421, $0.93) trails GPT-6 Luna max (1432, $0.07)"
  },
  {
   "task": "People, messages, money, records",
   "best": "GPT-6.1 Sol high ($0.32, AA-Omniscience 41). Sol low ($0.13, 38) for a cheaper first pass with verification",
   "avoid": "Haiku 5.5 (3 to 11) and Sonnet 5.5 (19 to 32) at any effort. Grok (26 to 32) and Gemini 3.8 Flash (29 to 30) at any effort trail GPT-6.1 Sol low (38, $0.13) Argon high costs $1.99 per AA task through November 30; estimated $3.98 from December 1 assuming unchanged tokens. Subscription caps and availability are separate."
  },
  {
   "task": "Computer use and browser work",
   "best": "Haiku 5.5 xhigh (OSWorld 67.5% at $0.30 per attempt). Only Anthropic and Luna are on this chart",
   "avoid": "Sonnet 5.5 max (+2.9 points for 2.5x). Haiku for steps that type names, accounts, or payments"
  },
  {
   "task": "Agentic coding",
   "best": "GPT-6.1 Sol xhigh (Terminal-Bench 54%, $0.39), ahead of Opus 5.5 medium (53%) at 29% of its cost",
   "avoid": "Opus 5.5 xhigh (+3 for 1.9x) and max (+0). Sonnet 5.5 max ($5.46). Grok at any effort (3 to 26%): Grok 4.7 xhigh ($3.74, 26%) trails GPT-6.1 Sol low ($0.13, 31%). Gemini 3.8 Flash scores 20% at medium and high. With Gemini, use Flash medium for loops, refactors, and hard bugs, since high adds 33% to the cost for no gain. Argon high costs $1.99 per AA task through November 30; estimated $3.98 from December 1 assuming unchanged tokens. Subscription caps and availability are separate."
  },
  {
   "task": "Small, clear coding fixes",
   "best": "GPT-6.1 Sol low ($0.13, Terminal-Bench 31%). Haiku 5.5 max ($0.21, 33%) in Claude",
   "avoid": "Opus 5.5 low for this work, since Haiku 5.5 max scores higher for 38% of the cost. Grok 4.7 low ($1.25, 16%) and Gemini 3.8 Flash medium ($0.93, 20%) trail Haiku 5.5 xhigh ($0.12, 29%)"
  },
  {
   "task": "Research and reasoning",
   "best": "GPT-6.1 Sol medium (HLE 50%, $0.21). Flash high (48%, $1.24) beats Flash medium (42%). Argon high is newly measured at 57% for $1.99 introductory; verify access before adopting it",
   "avoid": "Opus 5.5 xhigh and max: 2 to 3 points per cost doubling. Grok tops out at 44% (Grok 4.6 xhigh, $1.75), below GPT-6.1 Sol low (47%, $0.13) Argon high costs $1.99 per AA task through November 30; estimated $3.98 from December 1 assuming unchanged tokens. Subscription caps and availability are separate."
  },
  {
   "task": "General agentic judgment",
   "best": "GPT-6.1 Sol medium (index 48, $0.21)",
   "avoid": "Opus 5.5 xhigh and max (+2 per doubling). Sonnet 5.5 high (47, $0.88) trails Opus 5.5 medium on every test. Grok 4.7 high and xhigh (46) cost 13 to 18x GPT-6.1 Sol medium for a lower score. Gemini 3.8 Flash high adds 1 point over medium (41 vs 40) for 33% more. Fable 5.1 high (51, $3.91) ties Opus 5.5 medium (51, $1.34) at nearly 3x the task cost. Argon high costs $1.99 per AA task through November 30; estimated $3.98 from December 1 assuming unchanged tokens. Subscription caps and availability are separate."
  }
 ]
}