Will the next Claude Haiku model (4.6+) debut at a score of at least 1440 by June 30, 2027?

- [2026-10-07] Anthropic released Claude Haiku 5.5. It qualifies by name under the rules, because "Haiku 5.5" is a version above 4.6. The Anthropic homepage lists it as a launch dated Oct 7, 2026. Sources: https://www.anthropic.com/ and https://releasebot.io/updates/anthropic/claude - [2026-10-07] Arena posted on X that Haiku 5.5 was "now in the Arena" in Agent Arena and Code Arena (WebDev, Text, Document, Vision) and that scores were "coming soon". That means it was open for voting but had no published score. Source: https://x.com/arena/status/2107906881276826082. The X page itself could not be fetched, so this comes from the search snippet. - [2026-10-14] Crypto Briefing, fetched in full, says Haiku 5.5 (High) ranks #27 in Agent Arena and #30 in Code Arena WebDev. It gives no Text Arena score, and Text Arena is the board that resolves this market. Source: https://cryptobriefing.com/claude-haiku-5-5-agent-arena-ranking/ - [2026-10-15] I rendered the resolution page, dated Oct 15, 2026 (416 models, 8.83M votes). Ranks 1–57 contain no Haiku model. The scores around ranks 41–57 were 1458 down to 1448. Rank 30 was Claude Sonnet 5.5 (xHigh) at 1470 with only 6,831 votes. The rest of the table (rank 58 onward) was truncated, so I could not confirm whether Haiku 5.5 or 4.5 appears lower down. Source: https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-17] Searches turned up no Text Arena (no-style-control) score for Haiku 5.5. As of the Oct 15 snapshot it appears not yet added to the board, which is unconfirmed below rank 57. If it is added in the coming days, the market would resolve off that first-appearance score. Source: https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-15] Reference scores from the same board, for calibration against the 1440 strike: - Claude Sonnet 4.6 scored 1458. - Claude Opus 4.8 scored 1454. - Claude Opus 4.5 scored 1451. - Claude Sonnet 5.5 (xHigh) scored 1470, as a recent preliminary-sized entry. - Gemini 3 Flash scored 1466. - GLM-5.3-Flash scored 1471. - MiMo-V2.6-Flash scored 1455. - Cheap or small models from other labs sit in the 1455–1470 range. Source: https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-07] Anthropic pitches Haiku 5.5 as its cheapest, fastest and most capable small model, at $0.10 / $0.50 per million tokens, with a 1M-token context window. It is around 75% cheaper than Haiku 4.5. Source: https://cryptobriefing.com/claude-haiku-5-5-agent-arena-ranking/ - [2026-10] Third-party benchmarks put Haiku 5.5 well above Haiku 4.5. The Artificial Analysis Intelligence Index is 43 (18th), and benchlm shows 70.1/100 against 41.2 for Haiku 4.5. These are not Arena Elo and are only indirect evidence. Sources: https://tech-insider.org/claude-haiku-5-5-launch-pricing-2026/ and https://benchlm.ai/models/claude-haiku-5-5 - [2026-10-15] The rest of the Claude 5.5 family is already on the Text board: Opus 5.5 (High) at 1516, Fable 5.1 (Max) at 1512, and Sonnet 5.5 (xHigh) at 1470. Haiku 4.5 has not been replaced by a 4.6 model, and a Haiku 4.6 never shipped. Sources: https://arena.ai/leaderboard/text/overall-no-style-control and https://hidekazu-konishi.com/entry/anthropic_claude_model_release_timeline.html - [2026-10-07] Earlier Polymarket sibling markets on the Haiku release date, on a snippet dated about Oct 5, priced the release as near-certain by Dec 31. That release has since happened (Oct 7). Source: https://polymarket.com/event/next-claude-haiku-released-byptptpt-20260701205353326 - [2026-10-17] No scheduled events were found before the June 30, 2027 close beyond Arena publishing the Haiku 5.5 Text score. Any later Haiku (for example 5.6 or 6) would only matter if Haiku 5.5 is never added to Text Arena. BOTTOM LINE: Haiku 5.5 is the likely qualifying model. It launched Oct 7 and was in Arena voting, but no Text Arena score had appeared on the Oct 15 board (visible through rank 57). If it is added, a small-model score near or above 1440 looks plausible, given that recent comparable models sit around 1450–1470. The main uncertainties are the exact debut score (the score displayed the day after it first appears) and whether it was already added before the market opened, which would exclude it. NOT FOUND: A Haiku 5.5 score on the Text Arena (no-style-control) board. The board below rank 57 (truncated). Arena's own announcement of a Text score for Haiku 5.5. A Haiku score in the Arena changelog. Any Haiku 4.6 model.

- [2026-10-17] Anthropic released Claude Haiku 5.5 on 2026-10-07 (model ID claude-haiku-5-5). Its version number is above 4.6, so it would qualify. The release is confirmed. One tracker says there was never a Haiku 4.6 to 4.8 or a Haiku 5. https://www.anthropic.com/claude-haiku-5-5 and https://hidekazu-konishi.com/entry/anthropic_claude_model_release_timeline.html - [2026-10-17] Haiku 5.5 does not appear in the top 57 rows of the arena.ai no-style-control text leaderboard (data dated 2026-10-15, 416 models). I could not see the omitted lower rows of that snapshot. The 2026-10-08 snapshot lists claude-haiku-4-5-20251001 at 1396 and shows no Haiku 5.5 in its first ~212 rows. So the event has probably not resolved yet. The first appearance is the trigger, and it could land at any point from now to June 2027. https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-08] Haiku 4.5 sits at 1396 (±2, about 146k votes) on this leaderboard. The 1440 threshold is about 44 points above it. https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-15] Anthropic's larger models on the current board: Opus 5.5 (High) 1516, Fable 5.1 1512, Opus 5 (High) 1503, Sonnet 5.5 (xHigh) 1470 (±8, only 6.8k votes, so still settling), Sonnet 5 (High) 1442, Sonnet 4.6 1458, Sonnet 4.5 1439. https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-08] Reference class for the Haiku-to-Sonnet gap: Haiku 4.5 (1396) trailed Sonnet 4.5 (1439) by about 43 points. Applying a similar gap to Sonnet 5.5 (~1470) gives roughly 1425–1435 for Haiku 5.5, just below 1440. This is my arithmetic, not a published forecast. https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-08] Comparable small or cheap models from other labs: GPT-5.6 Luna (xHigh) 1430, GPT-6 Luna (Max) 1391, Gemini 3.5 Flash-Lite 1435, Gemini 3.5 Flash (High) 1482, Gemini 3.8 Flash (High) 1499 (preliminary), GLM-5.3-Flash 1471, GPT-5.4 mini 1411. Cheap-tier models scoring 1440 or more are now common, but not universal. https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-07] Anthropic's launch benchmarks say Haiku 5.5 beats Haiku 4.5 and GPT-6 Luna on every row and trails Sonnet 5.5 on every row. Press coverage says it "approaches" Sonnet 5.5 on some benchmarks. These are coding and agentic benchmarks, not arena preference scores. https://www.androidauthority.com/claude-haiku-5-5-launched-3720550/ and https://cellcog.ai/blog/claude-haiku-5-5-release-date/ - [2026-10-07] The model is priced at $0.10 input and $0.50 output per million tokens, with adaptive thinking and an effort setting. The arena may list effort variants such as High or Max separately, and the highest-scoring model added on the first day counts. That raises the chance that at least one variant clears 1440. The leaderboard already lists effort-labelled entries for Opus 5 and Sonnet 5.5. https://venturebeat.com/technology/anthropic-launches-claude-haiku-5-5-with-90-api-price-reduction-matching-gpt-6-luna and https://arena.ai/leaderboard/text/overall-no-style-control - [2026-02-20] Past lag from Anthropic release to arena listing: Sonnet 4.6 was released 2026-02-17 and the arena reported it at 1457 (#13) on 2026-02-20, about 3 days later. Opus 4.6 was released 2026-02-05 and had a score by late February. The first-appearance score can later drift, since Sonnet 4.6 now shows 1458. https://x.com/arena/status/2024883614249615394 - [2026-10-17] Haiku 5.5 was released 10 days ago and may already be on the board in rows I could not see. Sonnet 5.5 (released 2026-09-28) was on the 2026-10-08 board with 4.8k votes. If Haiku 5.5 is added after 2026-10-17, the question is live. If it was already added before the market opened and the question considers it, resolution is murkier. I could not verify the exact date it was added. - [2026-10-15] Preliminary labels: new models such as Gemini 4 Argon and Gemini 3.8 Flash carry a "Preliminary" tag, but they still show a score. The "AutoEval" exclusion in the rules is a separate label, and I saw no AutoEval labels in the visible rows. https://arena.ai/leaderboard/text/overall-no-style-control - [2026-10-17] Haiku generations are rare. They arrived 2024-03 (Claude 3 Haiku), 2024-10/11 (3.5), 2025-10 (4.5) and 2026-10 (5.5). This implies no further new Haiku model is likely before June 2027. The question therefore likely rests on Haiku 5.5's arena debut. https://hidekazu-konishi.com/entry/anthropic_claude_model_release_timeline.html BOTTOM LINE: The question almost certainly depends on Haiku 5.5, which is released but, as far as I can see, not yet in the visible part of the board. A Sonnet-minus-~40 analogy and peer small models point to a debut near 1425–1445, so 1440 is close to a coin-flip. Effort variants, where the highest-scoring one counts, tilt slightly toward Yes, and the possibility that it was already listed before the window tilts toward No. NOT FOUND: A confirmed arena.ai listing or score for Haiku 5.5 (the board rows beyond about 212 were truncated, and the 2026-10-15 snapshot showed only the top 57). The exact arena first-appearance date. Any prediction-market odds or expert forecast on the Haiku 5.5 score (I deliberately did not look). Any Anthropic statement on the next Haiku after 5.5.

YES means: Claude Haiku 5.5, or another new Haiku at version 4.6 or higher, first appears on the Arena Text (no-style-control) leaderboard without an AutoEval label, and its displayed Score at 12:00 PM ET the next day is at least 1440 (June 30, 2027 deadline). - [2026-10-16] The Arena changelog, fetched in full, lists "Claude Haiku 5.5 (High)" as added on Oct 14, but only to the Agent Arena leaderboard. No Text Arena entry for any Haiku appears through Oct 16. The Text board therefore has not yet had a Haiku 5.5 debut, so the market is still live. This is the strongest procedural point for YES: the qualifying event is still ahead of us. https://arena.ai/company/leaderboard-changelog - [2026-10-02] Sonnet 5.5 shipped Sep 28 and its xhigh variant reached Text Arena on Oct 2, about 4 days later. Arena also named Haiku 5.5 for Text on Oct 7, with "scores coming soon", per the Latent Space recap. Haiku 5.5's Agent Arena entry came Oct 14, a 7-day lag. A Text entry is plausibly next, since Arena has said it is testing Haiku 5.5 in Text. https://arena.ai/company/leaderboard-changelog and https://www.latent.space/p/ainews-claude-haiku-55-better-than - [2026-10-13] A blogger tracking cheap models on Arena wrote on Oct 13 that Haiku 5.5 had no Arena rating yet, and expected one "next week". They are not a primary source. https://www.stephanmiller.com/model-buzz-roundup-week-of-1007/ - [2026-10-13] Cheap-tier peers on the Text board, using the Oct 8 dataset as relayed by the same blog. These are style-controlled figures, so they may differ from the no-style-control board. GLM-5.3-Flash was at about 1475 with 17k+ votes. MiMo-V2.6-Pro was 1480. Opus 5.5 (High) was 1507. The cheapest models close to the frontier now cluster far above 1440. That sets the competitive bar Haiku 5.5 is being rated against. https://www.stephanmiller.com/model-buzz-roundup-week-of-1007/ - [2026-10-08] Artificial Analysis ranks Haiku 5.5 at 43, ahead of GLM-5.3 Flash (42), Gemini 3.8 Flash (41) and GPT-6 Luna (38). GLM-5.3-Flash sits at roughly 1471–1475 on Text. Gemini 3.8 Flash sits at about 1499 (a colleague's figure). If Arena preference tracks the Artificial Analysis ordering, Haiku 5.5 should land at or above GLM, which is well above 1440. The weakness of this link is that the Intelligence Index measures reasoning and agentic work, and Arena measures chat preference. https://artificialanalysis.ai/articles/claude-haiku-5-5 - [2026-10-12] Design Arena's Oct 12 post says Haiku 5.5 is up 203 Elo on Haiku 4.5 in Game Dev. It also says Haiku 5.5 took 1st in a website tournament, ahead of GPT-5.6 Sol (Medium), which costs 56x more. Design Arena shows an Elo of 1349 for Haiku 5.5 on its own scale. This shows large human-preference gains in a vote-based arena, though it is a different arena and scale. https://x.com/DesignArena/status/2109688565231993182 and https://www.designarena.ai/models - [2026-10-14] Code Arena WebDev data from Oct 13: Haiku 5.5 scores 1585 (±12, 2,771 votes), against GPT-6 Luna at 1582 (±6, 17,869 votes). That is a tie with a peer that sits around 1391–1430 on Text. Haiku 5.5 is competitive with the strongest cheap models in a human-vote arena, and its early vote count is low. https://www.gptunnel.ru/en/blog/haiku-5-5-vs-gpt-6-luna - [2026-10-14] The counter-signal, for fairness: on Agent Arena, Luna (max) is above Haiku 5.5 (High), ranked 19th against 27th. Haiku was run at High effort and Luna at Max. If Arena lists several Claude effort variants on Text, the rules take the highest-scoring one added on the first day. A Max or xhigh Haiku variant would help YES. https://www.gptunnel.ru/en/blog/haiku-5-5-vs-gpt-6-luna - [2026-10-07] Haiku 5.5 added effort levels (low to max) and adaptive thinking, plus a June 2026 knowledge cutoff and a new Sonnet 5.5-family tokenizer. The tokenizer and effort controls are the same architecture as Sonnet 5.5, which is at ~1470. Haiku 4.5 had none of these features. It scored 1396 on a model that predated them. That 1396 is the wrong anchor for a model with a one-year-newer training pipeline. https://computingforgeeks.com/claude-haiku-5-5-released-features-benchmarks/ - [2026-10-07] Closing the gap to Sonnet: third-party analysis says that on the two knowledge-work Elo benchmarks, Haiku 5.5 closes about 80% of the gap between Haiku 4.5 and Sonnet 5.5. Haiku 5.5 beats GPT-6 Luna on every row in Anthropic's launch table. That is a much smaller Haiku-to-Sonnet gap than the 43 points seen at 4.5, and a smaller gap points to a score above the colleagues' 1425–1435 estimate. https://computingforgeeks.com/claude-haiku-5-5-released-features-benchmarks/ - [2026-10-07] Anthropic's Haiku 5.5 over-refusal bug is documented. It depressed an Artificial Analysis automation score, and a fix is in progress. A safety-refusal regression is the main thing that could hurt a chat-preference score. If it is fixed before the Text debut, that risk shrinks. https://artificialanalysis.ai/articles/claude-haiku-5-5 BOTTOM LINE: The most likely path to YES is Arena adding Haiku 5.5 (a High, xhigh or Max variant) to Text in the next days or weeks, with its first-day score at or above 1440. That needs the Sonnet-5.5-class training and the 43 Artificial Analysis score to translate into a preference score close to GLM-5.3-Flash (~1471–1475) rather than 4.5-era ~1396, with the highest effort variant counting. It also needs no first appearance before Oct 1 (none found) and no AutoEval label. NOT FOUND: Any Haiku 5.5 score on the Text no-style-control board. A Text Arena changelog entry for Haiku 5.5 (the last changelog I read was Oct 16). Any Arena statement of a Text release date.

YES means: Haiku 5.5, or a later Haiku at version 4.6 or higher, first appears on the Arena Text (no-style-control) board without an AutoEval label, and its displayed Score at 12:00 PM ET the next day is 1440 or more. - [2026-10-16] The Arena changelog, fetched in full, lists Haiku 5.5 only as "added to the Agent Arena leaderboard" on Oct 14. It lists no Text Arena entry through Oct 16. Arena's Oct 7 post said Haiku was in Text for voting, but the changelog treats listing as a separate event. Anthropic's Sonnet 5.5 (xHigh) was added to Text on Oct 2, five days after its release. Haiku 5.5 is now 10 days past release with no Text entry. That gap is mild evidence of a slower or lower-priority Text listing, or of listing only after votes accumulate. It is not evidence the model will never be listed. https://arena.ai/company/leaderboard-changelog - [2026-10-16] The only Anthropic variant Arena has listed anywhere is "Haiku 5.5 (High)", and only in Agent Arena. Arena has not listed a Max or xHigh Haiku. Sonnet 5.5 debuted on Text as xHigh. If Haiku debuts on Text at High or Medium effort, a lower-effort first entry would score less than a Max entry. This weakens the "highest-scoring variant counts" tilt toward Yes in the earlier reports. https://arena.ai/company/leaderboard-changelog - [2026-10-15] I re-rendered the Text board. It is dated Oct 15, with 416 models and 8.83M votes, and no Haiku appears in ranks 1–57. The board's first-day scores for new entries are often noisy and preliminary. Gemini 4 Argon at 1532 has ±9 on 4.9k votes, and Sonnet 5.5 has ±8 on 6.8k votes. A Manifold note from Oct 2 says a new model's score usually drifts down as the early enthusiasts are joined by everyone else. Resolution uses the first-day number, so a noisy first score could land on either side of 1440. https://arena.ai/leaderboard/text/overall-no-style-control and https://manifold.markets/Terminator2/which-company-will-have-the-1-model - [2026-10-15] Cheap-tier peers on the same board are not uniformly above 1440. The earlier reports put Haiku 4.5 at 1396, GPT-6 Luna at 1391 and GPT-5.6 Luna at 1430. The Haiku 5.5 launch pitch compares it to GPT-6 Luna, which is Haiku's direct price match and sits about 50 points below the strike. Artificial Analysis places Haiku 5.5 at 43 versus 56 for Sonnet 5.5. That 13-point gap is large. Anthropic's own table also shows Sonnet 5.5 ahead on every row. If Sonnet's preliminary 1470 drifts down, a Haiku-minus-30-to-40 gap would put Haiku near 1430. https://artificialanalysis.ai/articles/claude-haiku-5-5 - [2026-10-07] Haiku 5.5's benchmark gains come from agentic and knowledge-work tests at max effort. Haiku 5.5's GDPval-AA score falls from 1620 at max effort to 1277 at the default medium effort, per the Spanish-language DEV writeup, which cites the launch table. Arena Text voters see default-style chat output. The over-refusal issue during pre-release testing (Artificial Analysis) and the high token use are both risks for human preference scores. An Arena Text score near 1440 would still be well below what the agentic numbers suggest. https://dev.to/roobia/evaluaciones-comparativas-de-claude-haiku-55-5h5e and https://artificialanalysis.ai/articles/claude-haiku-5-5 - [2026-10-14] Haiku 5.5 (High) is #27 in Agent Arena and #30 in Code Arena WebDev, per Crypto Briefing. The WebDev board is the only direct preference signal available so far. Rank #30 on WebDev is middling for a model that must reach about rank 60 on Text to clear 1440. Several rows around ranks 44–57 on the Text board sit at 1448–1455, so 1440 falls somewhere past rank 60. I could not verify that rank. https://cryptobriefing.com/claude-haiku-5-5-agent-arena-ranking/ - [2026-10-09] A third-party article stated that Haiku 5.5 was "not yet listed on the LMArena text leaderboard on 9 October 2026". That is consistent with the changelog and shows no earlier hidden listing. That means the "already listed before the window" risk is smaller than the earlier reports feared. The market stays live, but it hinges on a single not-yet-seen score. https://aiwiz.uk/blog/claude-haiku-5-5/ - [2026-10-17] Resolution technicalities that favour NO. (a) If Arena lists Haiku 5.5 only with an "AutoEval" label, or only on Agent, WebDev or Vision boards, it does not count. (b) If a Text entry appears and is then removed before 12:00 PM ET the next day, it does not count. (c) If the model is first listed in Text before the question's window was effectively available, ambiguity cuts toward the market's own date logic. (d) With Haiku generations about a year apart, no other qualifying model is likely before June 30, 2027. If Haiku 5.5 never lands on Text, the market resolves No. The earlier reports show Arena listing nearly every Anthropic launch on Text, so (d) is a tail risk. https://arena.ai/leaderboard/text/overall-no-style-control BOTTOM LINE: The strongest No path is a first-day Text score in the 1420s–1430s, from the Haiku-below-Sonnet gap, Luna-class peers and a debut at lower effort than Max. A smaller No path is delayed or absent Text listing; the changelog shows Haiku 5.5 only in Agent Arena as of Oct 16. Neither path looks decisive, because GLM-5.3-Flash, Gemini 3.8 Flash and MiMo Flash all sit well above 1440 at similar Artificial Analysis levels. NOT FOUND: A Haiku 5.5 Text Arena score or listing date. Any Arena statement on when Text scores arrive. The Text board below rank 57 in the Oct 15 render. Any effort-variant naming for a Text entry.