GEM-4C 581 —

GPT-61S 577 —

CL-SN55 575 —

CL-OP55 574 —

GPT-61TERR 567 —

CL-OP55EH 564 —

GPT-61SOLXH 561 —

GPT-61SOLH 547 —

CL-SN55XH 542 —

CL-OP55H 527 —

GPT-61TERRXH 525 —

GEM-4AR 518 —

GPT-61SOLM 518 —

KIMI-K31P 517 —

GPT-61TERRH 506 —

CL-OP55M 501 —

MUSE-WONDFUL 498 —

CL-SN55H 483 —

GPT-61TERRM 477 —

CL-HK55 477 —

GPT-6A 473 —

MUSE-WOND 468 —

CL-HK55XH 464 —

GPT-6ASTRXH 462 —

GPT-6ASTRH 459 —

GROK-47H 451 —

DSK-V-340615 449 —

GROK-47 447 —

QWEN-38X 441 —

GPT-61SOLL 437 —

GPT-6ASTRM 435 —

GLM-53 433 —

MIMO-V26P 431 —

CL-HK55H 431 —

GEM-38F 428 —

LING-31F1019 428 —

MMAX-M31F 426 —

GEM-38FM 424 —

CL-FABL51XH 423 —

GPT-61TERRNO 423 —

GROK-47L 422 —

CL-FABL51 418 —

GPT-6L 418 —

GLM-53F 418 —

STEP-5 416 —

GPT-61TERRL 415 —

MIMO-V26F 411 —

CL-OP55L 407 —

CL-SN55M 407 —

CL-FABL51H 404 —

MIST-LARG4 401 —

CL-HK55M 395 —

GPT-6LUNAXH 387 —

MUSE-SPAROPE 385 —

MIST-LARG4M 380 —

GPT-6ASTRL 376 —

CL-FABL51M 376 —

DSK-V4P 374 —

GLM-53L 373 —

GPT-6LUNAH 373 —

CL-SN55L 372 —

GEM-38FL 370 —

QWEN-827B 357 —

CL-HK55L 354 —

MMAX-M3 352 —

GPT-6LUNAM 352 —

DSK-V-02811D 349 —

K2-HORI375BA 349 —

REFL-BEAM 345 —

CL-FABL51L 344 —

QWEN-827BM 338 —

QWEN-827BL 326 —

NEMO-35SUPEV 319 —

INKL 300 —

NVID-NEMO3UL 299 —

GEM-35FLITE 299 —

GPT-6LUNAL 291 —

GPT-6LUNANON 289 —

QWEN-827BNON 286 —

COMM-APLUS 269 —

MUSE-GLIM 266 —

NEMO-35LIGH 263 —

MIST-SMAL4 253 —

INKL-SM 241 —

GEMM-431B 225 —

GEMM-4E4B 225 —

GEMM-4E4BNON 225 —

LING-30TINY 225 —

MINI-33B 225 —

NORT-MCODE 225 —

NOVA-MICR 225 —

NVID--ADE3B4 225 —

QWEN-54B 225 —

QWEN-54BNONR 225 —

TINY-AYAGLOB 225 —

AGNE-25PBETA 224 —

APER-8BINST 224 —

GEMM-412B 224 —

GEMM-412BNON 224 —

GEMM--BB78FD 224 —

GEMM--2ACF33 224 —

GEMM-431BNON 224 —

GPT-OSS20B 224 —

GPT-OSS20BL 224 —

GRAN-428B 224 —

LING-30FVL 224 —

LLAM-32INST1 224 —

MINI-314B 224 —

MINI-38B 224 —

NVID--29A159 224 —

NVID--F51AF2 224 —

NVID--342C05 224 —

PHI-4M 224 —

PHI-4MULT 224 —

QUAS-438B 224 —

QWEN-59B 224 —

QWEN-59BNONR 224 —

QWEN-8FNEXT 224 —

REKA-F3 224 —

SARV-105B 224 —

SOLA-MINI 224 —

CELE-1 223 —

GPT-OSS120B 223 —

GPT-OSS120BL 223 —

HY3 223 —

LFM2-526B 223 —

MERC-25 223 —

MIST-SMAL4NO 223 —

NVID--2A922F 223 —

PHI-4 223 —

QWEN-5OMNIF 223 —

LING-31F 222 —

LLAM-33INST7 222 —

NEMO-3NOMNI3 222 —

QWEN-CODENEX 222 —

QWEN--193CA5 222 —

QWEN--FE4EEB 222 —

QWEN--7CEE23 222 —

DSK-V4FVISI 221 —

MAGI-SMAL250 220 —

QWEN-7PLUS 220 —

NOVA-20LITE 218 —

NOVA--2F7EB0 218 —

NOVA--D2FCC4 218 —

NOVA--C277BF 218 —

QWEN--32D621 218 —

RING-261T 218 —

APOD-11 216 —

HERM--B5E53D 216 —

HERM--93BBC7 216 —

LONG-20 216 —

QWEN--43BC51 216 —

QWEN--0C45C9 216 —

QWEN--50732D 214 —

QWEN--C9949D 214 —

DSK-V4PNONR 213 —

KIMI-K27CODE 213 —

QWEN-5OMNIPL 211 —

QWEN-824TA95 206 —

O3 201 —

NOVA-20P 196 —

NOVA-20PRL 196 —

NOVA-20PRM 196 —

COMM-A 195 —

GPT-6SOLDAYB 195 —

GPT-53CODE 184 —

GRAN-423B 183 —

NVID-NEMO3N4 163 —

SOLA-PRO4 151 —

A-XK2 150 —

APER-70BINST 150 —

APRI-V1615BT 150 —

COGI-V21R 150 —

DEEP-3LLAM31 150 —

DEEP-3MIST24 150 —

DIFF-26BA4B 150 —

DOUB-SEEDCOD 150 —

ERNI-45300BA 150 —

ERNI-50TPR 150 —

EXAO-4012B 150 —

EXAO-4012BR 150 —

EXAO-4032B 150 —

EXAO-4032BR 150 —

EXAO-4533B 150 —

EXAO-4533BNO 150 —

FALC-H1R7B 150 —

G9V3-39A5 150 —

G9V3-3B 150 —

GEM-3DEEPTHI 150 —

GEMM-3270M 150 —

GEMM-4E2B 150 —

GEMM-4E2BNON 150 —

GPT-55P 150 —

GRAN-40350M 150 —

GRAN-40H350M 150 —

GRAN-40HN1B 150 —

GRAN-40HSMAL 150 —

GRAN-40MICR 150 —

GRAN-4230B 150 —

HERM--D72525 150 —

HERM--616704 150 —

HYPE-XSEEDTH 150 —

HYPE-60B 150 —

INTE-3 150 —

JAMB-17LARG 150 —

JAMB-17M 150 —

JAMB-R3B 150 —

JT-35BF 150 —

JT-41-A362F0 150 —

JT-41-0CD309 150 —

JT-M 150 —

K-EXAO200803 150 —

K2-HORI09B 150 —

K2-HORI37B 150 —

K2-HORI7B 150 —

K2-HORIMOVA3 150 —

K2-THINV2 150 —

K2-V2 150 —

K2-V2L 150 —

K2-V2M 150 —

KAT-CODEPV2 150 —

KIMI-LINE48B 150 —

LFM2-24BA2B 150 —

LFM2-512BINS 150 —

LFM2-512BT 150 —

LFM2-58BA1B 150 —

LFM2-5VL16B 150 —

LING-30FFIN 150 —

LLAM-31INST4 150 —

LLAM-31NEMOU 150 —

LLAM-32INST9 150 —

LONG-FLITE 150 —

MAGI-M2509 150 —

MI-DMK25PDEC 150 —

MINI-V4613B 150 —

MINI-1B 150 —

MINI-1BNONR 150 —

MINI-2B 150 —

MOLM-7BD 150 —

MOLM-8B 150 —

MOTI-2127B 150 —

MOTI-3 150 —

NANB-13B 150 —

NEMO-35BOLT7 150 —

NEMO-35BOLTE 150 —

NEMO-CASC230 150 —

NEX-N2P 150 —

NOVA-20OMNI 150 —

NOVA--A0BE81 150 —

NOVA--C3EC3E 150 —

NOVA-PREM 150 —

NVID--3F35FB 150 —

OLMO-3132BIN 150 —

OLMO-3132BTH 150 —

OLMO-37BINST 150 —

OLMO-37BTHIN 150 —

QWEN-508B 150 —

QWEN-508BNON 150 —

QWEN-52B 150 —

QWEN-52BNONR 150 —

QWEN--B6BFD1 150 —

R1-1776 150 —

RING-F20 150 —

SARV-30B 150 —

SOLA-OPEN100 150 —

SOLA-OPEN250 150 —

STEP-3VL10B 150 —

TRI-21BTHINV 150 —

LLAM-4SCOU 148 —

LLAM-4MAVE 147 —

QWEN--58A201 145 —

GPT-55INST06 136 —

MIST-M35 129 —

GEM-31PPR 128 —

TRIN-LARGT 76 —

GEM-4C 581 —

GPT-61S 577 —

CL-SN55 575 —

CL-OP55 574 —

GPT-61TERR 567 —

CL-OP55EH 564 —

GPT-61SOLXH 561 —

GPT-61SOLH 547 —

CL-SN55XH 542 —

CL-OP55H 527 —

GPT-61TERRXH 525 —

GEM-4AR 518 —

GPT-61SOLM 518 —

KIMI-K31P 517 —

GPT-61TERRH 506 —

CL-OP55M 501 —

MUSE-WONDFUL 498 —

CL-SN55H 483 —

GPT-61TERRM 477 —

CL-HK55 477 —

GPT-6A 473 —

MUSE-WOND 468 —

CL-HK55XH 464 —

GPT-6ASTRXH 462 —

GPT-6ASTRH 459 —

GROK-47H 451 —

DSK-V-340615 449 —

GROK-47 447 —

QWEN-38X 441 —

GPT-61SOLL 437 —

GPT-6ASTRM 435 —

GLM-53 433 —

MIMO-V26P 431 —

CL-HK55H 431 —

GEM-38F 428 —

LING-31F1019 428 —

MMAX-M31F 426 —

GEM-38FM 424 —

CL-FABL51XH 423 —

GPT-61TERRNO 423 —

GROK-47L 422 —

CL-FABL51 418 —

GPT-6L 418 —

GLM-53F 418 —

STEP-5 416 —

GPT-61TERRL 415 —

MIMO-V26F 411 —

CL-OP55L 407 —

CL-SN55M 407 —

CL-FABL51H 404 —

MIST-LARG4 401 —

CL-HK55M 395 —

GPT-6LUNAXH 387 —

MUSE-SPAROPE 385 —

MIST-LARG4M 380 —

GPT-6ASTRL 376 —

CL-FABL51M 376 —

DSK-V4P 374 —

GLM-53L 373 —

GPT-6LUNAH 373 —

CL-SN55L 372 —

GEM-38FL 370 —

QWEN-827B 357 —

CL-HK55L 354 —

MMAX-M3 352 —

GPT-6LUNAM 352 —

DSK-V-02811D 349 —

K2-HORI375BA 349 —

REFL-BEAM 345 —

CL-FABL51L 344 —

QWEN-827BM 338 —

QWEN-827BL 326 —

NEMO-35SUPEV 319 —

INKL 300 —

NVID-NEMO3UL 299 —

GEM-35FLITE 299 —

GPT-6LUNAL 291 —

GPT-6LUNANON 289 —

QWEN-827BNON 286 —

COMM-APLUS 269 —

MUSE-GLIM 266 —

NEMO-35LIGH 263 —

MIST-SMAL4 253 —

INKL-SM 241 —

GEMM-431B 225 —

GEMM-4E4B 225 —

GEMM-4E4BNON 225 —

LING-30TINY 225 —

MINI-33B 225 —

NORT-MCODE 225 —

NOVA-MICR 225 —

NVID--ADE3B4 225 —

QWEN-54B 225 —

QWEN-54BNONR 225 —

TINY-AYAGLOB 225 —

AGNE-25PBETA 224 —

APER-8BINST 224 —

GEMM-412B 224 —

GEMM-412BNON 224 —

GEMM--BB78FD 224 —

GEMM--2ACF33 224 —

GEMM-431BNON 224 —

GPT-OSS20B 224 —

GPT-OSS20BL 224 —

GRAN-428B 224 —

LING-30FVL 224 —

LLAM-32INST1 224 —

MINI-314B 224 —

MINI-38B 224 —

NVID--29A159 224 —

NVID--F51AF2 224 —

NVID--342C05 224 —

PHI-4M 224 —

PHI-4MULT 224 —

QUAS-438B 224 —

QWEN-59B 224 —

QWEN-59BNONR 224 —

QWEN-8FNEXT 224 —

REKA-F3 224 —

SARV-105B 224 —

SOLA-MINI 224 —

CELE-1 223 —

GPT-OSS120B 223 —

GPT-OSS120BL 223 —

HY3 223 —

LFM2-526B 223 —

MERC-25 223 —

MIST-SMAL4NO 223 —

NVID--2A922F 223 —

PHI-4 223 —

QWEN-5OMNIF 223 —

LING-31F 222 —

LLAM-33INST7 222 —

NEMO-3NOMNI3 222 —

QWEN-CODENEX 222 —

QWEN--193CA5 222 —

QWEN--FE4EEB 222 —

QWEN--7CEE23 222 —

DSK-V4FVISI 221 —

MAGI-SMAL250 220 —

QWEN-7PLUS 220 —

NOVA-20LITE 218 —

NOVA--2F7EB0 218 —

NOVA--D2FCC4 218 —

NOVA--C277BF 218 —

QWEN--32D621 218 —

RING-261T 218 —

APOD-11 216 —

HERM--B5E53D 216 —

HERM--93BBC7 216 —

LONG-20 216 —

QWEN--43BC51 216 —

QWEN--0C45C9 216 —

QWEN--50732D 214 —

QWEN--C9949D 214 —

DSK-V4PNONR 213 —

KIMI-K27CODE 213 —

QWEN-5OMNIPL 211 —

QWEN-824TA95 206 —

O3 201 —

NOVA-20P 196 —

NOVA-20PRL 196 —

NOVA-20PRM 196 —

COMM-A 195 —

GPT-6SOLDAYB 195 —

GPT-53CODE 184 —

GRAN-423B 183 —

NVID-NEMO3N4 163 —

SOLA-PRO4 151 —

A-XK2 150 —

APER-70BINST 150 —

APRI-V1615BT 150 —

COGI-V21R 150 —

DEEP-3LLAM31 150 —

DEEP-3MIST24 150 —

DIFF-26BA4B 150 —

DOUB-SEEDCOD 150 —

ERNI-45300BA 150 —

ERNI-50TPR 150 —

EXAO-4012B 150 —

EXAO-4012BR 150 —

EXAO-4032B 150 —

EXAO-4032BR 150 —

EXAO-4533B 150 —

EXAO-4533BNO 150 —

FALC-H1R7B 150 —

G9V3-39A5 150 —

G9V3-3B 150 —

GEM-3DEEPTHI 150 —

GEMM-3270M 150 —

GEMM-4E2B 150 —

GEMM-4E2BNON 150 —

GPT-55P 150 —

GRAN-40350M 150 —

GRAN-40H350M 150 —

GRAN-40HN1B 150 —

GRAN-40HSMAL 150 —

GRAN-40MICR 150 —

GRAN-4230B 150 —

HERM--D72525 150 —

HERM--616704 150 —

HYPE-XSEEDTH 150 —

HYPE-60B 150 —

INTE-3 150 —

JAMB-17LARG 150 —

JAMB-17M 150 —

JAMB-R3B 150 —

JT-35BF 150 —

JT-41-A362F0 150 —

JT-41-0CD309 150 —

JT-M 150 —

K-EXAO200803 150 —

K2-HORI09B 150 —

K2-HORI37B 150 —

K2-HORI7B 150 —

K2-HORIMOVA3 150 —

K2-THINV2 150 —

K2-V2 150 —

K2-V2L 150 —

K2-V2M 150 —

KAT-CODEPV2 150 —

KIMI-LINE48B 150 —

LFM2-24BA2B 150 —

LFM2-512BINS 150 —

LFM2-512BT 150 —

LFM2-58BA1B 150 —

LFM2-5VL16B 150 —

LING-30FFIN 150 —

LLAM-31INST4 150 —

LLAM-31NEMOU 150 —

LLAM-32INST9 150 —

LONG-FLITE 150 —

MAGI-M2509 150 —

MI-DMK25PDEC 150 —

MINI-V4613B 150 —

MINI-1B 150 —

MINI-1BNONR 150 —

MINI-2B 150 —

MOLM-7BD 150 —

MOLM-8B 150 —

MOTI-2127B 150 —

MOTI-3 150 —

NANB-13B 150 —

NEMO-35BOLT7 150 —

NEMO-35BOLTE 150 —

NEMO-CASC230 150 —

NEX-N2P 150 —

NOVA-20OMNI 150 —

NOVA--A0BE81 150 —

NOVA--C3EC3E 150 —

NOVA-PREM 150 —

NVID--3F35FB 150 —

OLMO-3132BIN 150 —

OLMO-3132BTH 150 —

OLMO-37BINST 150 —

OLMO-37BTHIN 150 —

QWEN-508B 150 —

QWEN-508BNON 150 —

QWEN-52B 150 —

QWEN-52BNONR 150 —

QWEN--B6BFD1 150 —

R1-1776 150 —

RING-F20 150 —

SARV-30B 150 —

SOLA-OPEN100 150 —

SOLA-OPEN250 150 —

STEP-3VL10B 150 —

TRI-21BTHINV 150 —

LLAM-4SCOU 148 —

LLAM-4MAVE 147 —

QWEN--58A201 145 —

GPT-55INST06 136 —

MIST-M35 129 —

GEM-31PPR 128 —

TRIN-LARGT 76 —

Some supplementary benchmark or Arena observations are delayed.

← Back to feed

Sep 24, 2026

Anthropic Ran a Two-Week Sprint Where Claude Benchmarked and Fixed Its Own App to a 3.1x Speedup

Anthropic ran a two-week performance sprint on claude.ai’s web and desktop apps in August 2026. Claude built the measurement benchmarks, investigated bottlenecks, and authored approximately 3,000 code changes across the sprint. Engineers selected the performance targets, decided what to measure, and reviewed and approved deployments before they went live. The average speedup across 13 measured interactions was 3.1x.

Division of work

The sprint split responsibilities clearly:

Engineers: chose which journeys to target, defined what to measure, reviewed and approved every change before deployment
Claude: built the benchmarks, investigated the root causes of slowdowns, and submitted the code changes — roughly 3,000 of them over two weeks

Anthropic’s framing from the write-up: with Claude, measuring something makes it tractable. Setting up deterministic benchmarks was the enabling step; once measurement existed, Claude could work through the bottleneck list.

The numbers

Thirteen specific user interactions in the claude.ai web and desktop apps were measured before and after the sprint. The average speedup across those 13 interactions was 3.1x, calculated as the geometric mean of p75 timings. Published per-interaction examples:

Fresh page load (claude.ai): 3,085 ms → 550 ms
Message sending: 928 ms → 48 ms

The sprint covered both the web client and the desktop application.

What makes this distinct

Claude writing code is not new — Anthropic reported in June that Claude authors more than 80% of its own production code. The sprint result illustrates a different configuration of that capability: Claude also built the measurement infrastructure that told it what to fix. The code-writing and the benchmark-writing came from the same system.

The practical implication is that performance engineering at this scale — 3,000 changes in two weeks, across two clients — becomes feasible with a loop where the agent both measures and fixes, rather than one where engineers instrument and agents only implement. Engineers remain the judgment layer on what to target and what to ship; the analytical and implementation work scales under that oversight.